Skip to main content

Model family by Alibaba

Qwen

Alibaba's Qwen family of open-weight and hosted models.

Where it sits in Alibaba's lineup

  1. QwenAlibaba's Qwen family of open-weight and hosted models.This model
Made by
Alibaba
Releases
2 tracked
Current release
Qwen3.8 27B

Strengths and weaknesses

Compare with other models Up to 3 beside Qwen3.8 27B

  • Qwen3.8 27B

Suggested comparisons

Qwen3.8 27B is the newest Qwen release. Add its earlier releases to see how the line moved.

Qwen3.8 27BAlibabaThis page, always on the chart

Among the 40 current releases on this chart, Qwen3.8 27B's best rank is 18th of 38 on Long context. Agentic coding is Artificial Analysis's run on Terminal-Bench 4.0, Science is Artificial Analysis's run on GPQA Diamond, Computer use is Alibaba's own figure on OSWorld-Verified and Instruction following is Alibaba's own figure on IFBench, marked on the chart. It ranks lowest on Factuality, 30th of 30 and Hard reasoning, 35th of 38. Math: not yet published by Alibaba or any independent evaluator, checked Nov 2, 2026. Answers in full: published only for another harness, setting or build, listed in the table.

Each axis is one public benchmark at one setting, on its own fixed 0 to 100 scale, so adding a model never moves another. A dotted bridge crosses a missing result, which is never a zero. A hollow mark is a result from another source: the maker's own figure or another lab's run, where the benchmark's usual source has not tested the model yet. The readout names it.

On the chart

  • Qwen3.8 27BThis pageResults on 10 of 12 axes$0.50 / $338 tokens/sFirst token in 4.1 s

Prices and speeds read between Oct 28, 2026 and Nov 2, 2026.

Qwen3.8 27B on each axis10 of 12 with a result

Beside the benchmarks

Cost, speed, capacity, quality and accuracy for Qwen3.8 27B, with where each figure sits among the 40 models GILD tracks. These measures stand beside the chart and never feed it.

Cost

$1.15per task

Cost per task. One Intelligence Index task at list price, every token counted. Run at extra-high effort.

Speed

42tokens a second

Output speed. How fast it writes once the answer starts. Run at extra-high effort.

First token
3.9 s
Whole answer
64 s

Capacity

256Ktokens

Context window. The most one request can hold, as the maker states it.

Max output
262K
Reads
Text, Images, Video

Quality

33.7of 100

Intelligence Index. One composite score across independent evaluations. Run at extra-high effort.

GDPval knowledge work
1,427 Elo
Terminal-Bench 4.0
5.6%

Hallucination

28.6%wrong when it should say “I don’t know”

Hallucination rate. Of the questions it can't get right, how often it answers anyway. Lower is better. Run at extra-high effort.

Knowledge accuracy
17.3%

Where you can get it

3of 19 routes

Routes offering it. Of the clouds, gateways and coding products GILD tracks, those offering it today.

Open weights
Yes
Licence
Apache 2.0
Every route and price

Figures from Artificial Analysis, OpenRouter and GILD's route matrix, read Oct 27 to Nov 2, 2026. Each measure's definition and source open under . Not yet published for Qwen3.8 27B: Cached input and Epoch Capabilities Index.

Every value with its benchmark, run and source

Qwen3.8 27B's result on each axis, with the benchmark, the setting and run, who reported it, the source and the date
AxisBenchmark and settingQwen3.8 27BRunSourceDate
Agentic codingTerminal-Bench 4.0Accuracy, Artificial Analysis run in its own agent harness. Artificial Analysis's run at xhigh effort in its own agent harness, from its Intelligence Index evaluations. Vals.ai has not run it.Stands in for Terminal-Bench 4.0, pass@1 averaged over 3 runs, mini-swe-agent harness, which has no run of Qwen3.8 27B yet.5.6%Unusual next to its other results. Artificial Analysis's run at xhigh effort in its own agent harness, from its Intelligence Index evaluations. Vals.ai has not run it. Artificial Analysis run of Nov 2, 2026.Qwen3.8 27B (Xhigh)Effort: xhigh. Harness: Artificial Analysis agent harness.Artificial AnalysisRun by Artificial AnalysisNov 2, 2026
CodeSciCodeSubproblem accuracy, pass@1, highest effort run46.6%Qwen3.8 27B (Xhigh)Effort: xhigh. Harness: none.Artificial AnalysisNov 2, 2026
MathFrontierMath Tiers 1-3Accuracy, pass@1, highest effort runNot yet published by Alibaba or any independent evaluator, checked Nov 2, 2026
ScienceGPQA DiamondAccuracy, pass@1, Artificial Analysis run. Artificial Analysis's run at xhigh effort, the highest it ran for this release, from its Intelligence Index evaluations. Epoch AI has not run it.Stands in for GPQA Diamond, accuracy, pass@1, highest effort run, which has no run of Qwen3.8 27B yet.90.5%Qwen3.8 27B (Xhigh)Effort: xhigh. Harness: none.Not drawn: Alibaba 89.2% in its maker's own harness, Aug 14, 2026. SourceArtificial AnalysisRun by Artificial AnalysisNov 2, 2026
Hard reasoningHumanity's Last ExamAccuracy, text-only questions, no tools33.9%Qwen3.8 27B (Xhigh)Effort: xhigh. Harness: none.Artificial AnalysisNov 2, 2026
Tool useAutomationBench-AAScore, Artificial Analysis run48.2%Qwen3.8 27B (Xhigh)Effort: xhigh. Harness: Artificial Analysis AutomationBench harness.Artificial AnalysisNov 2, 2026
Computer useOSWorld-VerifiedSuccess rate as Alibaba reports it, step budget not stated. Alibaba's own figure from the Qwen3.8 27B model card. It does not state the harness, step budget or reasoning effort.Stands in for OSWorld 2.1, partial score, full 108-task set, v2.1 release, which has no run of Qwen3.8 27B yet.84.3%Qwen3.8-27BEffort: not published. Harness: not stated by Qwen.Not drawn: Alibaba 64.8% in its maker's own harness, Aug 14, 2026. SourceNot drawn: Alibaba 81.9% in its maker's own harness, Aug 14, 2026. SourceAlibabaReported by AlibabaAug 14, 2026
Long contextAA-LCR v1.1Accuracy, Artificial Analysis run82%Qwen3.8 27B (Xhigh)Effort: xhigh. Harness: none.Artificial AnalysisNov 2, 2026
MultimodalMMMU-ProOverall accuracy, no tools76.3%Qwen3.8 27B (Xhigh)Effort: xhigh. Harness: none.Artificial AnalysisNov 2, 2026
FactualityAA-Omniscience AccuracyAccuracy, Artificial Analysis run17.3%Hallucination rate 28.6% in the same run (lower is better)Qwen3.8 27B (Xhigh)Effort: xhigh. Harness: none.Artificial AnalysisNov 2, 2026
Answers in fullSpeechMap% Complete over 2,120 contested requests, reasoning-on entryPublished only for another harness, setting or build, listed in the tableNot drawn: SpeechMap.AI 46.3% at another setting, Oct 28, 2026. Source
Instruction followingIFBenchAccuracy as Alibaba reports it, effort not stated. Alibaba's own figure from the Qwen3.8 27B model card. It does not state the reasoning effort. Scale AI has not run MultiChallenge on it.Stands in for MultiChallenge, accuracy, Scale AI run, which has no run of Qwen3.8 27B yet.79.5%Qwen3.8-27BEffort: not published. Harness: not stated by Qwen.AlibabaReported by AlibabaAug 14, 2026

Price and speed

  • Qwen3.8 27B lists at $0.50 per million input tokens and $3 per million output tokens. Alibaba Cloud pricing, Oct 28, 2026.
  • Qwen3.8 27B, at extra-high effort, writes 38.3 output tokens per second on Alibaba's API, and its first token arrives 4.1 s after the request. Artificial Analysis, Nov 2, 2026.
  • Fastest on the chart is the shortest time to a 500-token answer: time to first token plus 500 tokens at the median output speed.

Every model with a result, axis by axis

57 models. A maker's own figure and another lab's run are marked with their benchmark; a missing result reads "Not yet published", never zero.

The result of every model GILD has benchmark results for, on each of the 12 axes. A vendor-reported value is marked. A model's run notes open under its row from its name, each with the axes it covers.
ModelAgentic codingCodeMathScienceHard reasoningTool useComputer useLong contextMultimodalFactualityAnswers in fullInstructions
Qwen3.8 27B5.6%Run by Artificial AnalysisTerminal-Bench 4.046.6%Not yet published90.5%Run by Artificial AnalysisGPQA Diamond33.9%48.2%84.3%Reported by AlibabaOSWorld-Verified82%76.3%17.3%Not yet published79.5%Reported by AlibabaIFBench
Qwen3.8 Max38.9%52.1%65.6%92.3%43.1%56.2%Not yet published80.3%82.8%29.5%65.3%Not yet published
Claude Haiku 5.536.4%55%75.1%89.6%44.4%35.4%72.4%Reported by AnthropicOSWorld 2.182.7%Not yet published34.7%78.8%67.7%
Claude Sonnet 5.564.1%61%88.8%95.6%55%71.8%80.1%Reported by Anthropic82.7%Not yet published51.5%81.6%Not yet published
Claude Opus 5.565.2%66.9%91.2%90.6%61.4%69.5%81.8%Reported by Anthropic84.7%87.7%65.5%82.8%Not yet published
Claude Fable 5.158.1%63.1%90.2%93.4%Run by Vals.ai59.1%59.4%80.7%Reported by Anthropic85.3%86.3%Run by Moonshot AI68.4%72.4%Not yet published
Claude Opus 553.5%56.4%85.6%93.9%54.9%56.6%77.7%79.3%84.7%60.9%73.7%Not yet published
Claude Sonnet 510.6%54.3%65.6%80.3%41.3%36.5%57%Reported by Anthropic82%77.3%40.1%Not yet publishedNot yet published
Claude Fable 541.4%61%87%85.9%55.5%54.1%Not yet published82.3%Not yet published65.4%64.8%63.5%Run by Artificial AnalysisIFBench
Claude Opus 4.823.2%54.4%80%91%48.7%45.6%54.8%Run by OSWorld teamOSWorld 2.077.7%Not yet published48.8%67.4%62.2%Run by Artificial AnalysisIFBench
Claude Opus 4.7Not yet published54.5%70.2%86.4%42.3%Not yet published48.9%Run by OSWorld teamOSWorld 2.078.7%78.8%48.9%70.2%58.6%Run by Artificial AnalysisIFBench
Claude Sonnet 4.63%Run by Artificial AnalysisTerminal-Bench 4.050.1%71.1%Run by Epoch AIOTIS Mock AIME 2024-202578.8%33.6%20.1%41.5%Run by OSWorld teamOSWorld 2.080%73.3%40.9%35.8%56.6%Run by Artificial AnalysisIFBench
Claude Opus 4.6Not yet publishedNot yet published66%88.4%39.9%Not yet publishedNot yet published78%75.4%47%59.9%37.2%
Claude Opus 4.5Not yet publishedNot yet published34.4%86%30.1%Not yet publishedNot yet published77.3%74%46.6%57.5%59%
Claude Haiku 4.5Not yet published42.2%66.7%Run by Epoch AIOTIS Mock AIME 2024-202571.2%10.4%3.2%Not yet published74.3%58.6%18%24.7%50.5%
Claude Sonnet 4.5Not yet published45.7%23.9%82.3%17.8%14.1%Not yet published72.3%68.7%32.9%31.4%55.3%
GPT-6.1 Terra49.5%55.1%90.9%94.3%46.2%60.2%74.2%Reported by OpenAIOSWorld 2.182.7%84.7%47.6%Not yet published74.2%
GPT-6.1 Sol55.1%54.2%93.7%95.4%52.9%64.9%78%Reported by OpenAIOSWorld 2.183%86%61.8%69.5%82%
GPT-6 Luna13.6%54.6%78.9%90.5%38.5%53.2%48.9%Run by AnthropicOSWorld 2.183.3%79.7%39.6%63.4%Not yet published
GPT-6 Astra59.6%56.5%93.7%95.8%54.7%68.5%75.8%Run by Meta80.7%86.9%61.6%68.2%Not yet published
GPT-5.6 Sol37.9%57.1%89.1%93.5%49.5%60.1%62.7%Run by OSWorld teamOSWorld 2.084%83.4%69.7%Run by Epoch AISimpleQA VerifiedNot yet published72.7%Run by Artificial AnalysisIFBench
GPT-5.6 Terra22.7%55%86%93.3%42.9%59.6%Not yet published83%80.7%42.3%Not yet published71.2%Run by Artificial AnalysisIFBench
GPT-5.6 Luna11.6%53.6%82.1%91.6%39.5%50.2%Not yet published83.7%78.6%41%Run by Epoch AISimpleQA VerifiedNot yet publishedNot yet published
GPT-5.514.6%Run by Artificial AnalysisTerminal-Bench 4.055.8%85.3%93.5%Run by Artificial Analysis45.8%47.3%49.5%Run by OSWorld teamOSWorld 2.084.3%79.9%63%Run by Epoch AISimpleQA VerifiedNot yet published75.9%Run by Artificial AnalysisIFBench
GPT-5.4Not yet publishedNot yet published78.6%93.3%43.7%Not yet publishedNot yet published82%78.4%Not yet publishedNot yet published73.9%Run by Artificial AnalysisIFBench
gpt-oss-120b14.2%Run by Terminal-Bench (via Epoch AI)Terminal-Bench 2.034%88.9%Run by Epoch AIOTIS Mock AIME 2024-202575.8%19.6%0.2%Not yet published52%Not yet published16.8%Reported by OpenAISimpleQA36.3%45.3%
GPT-6 Sol44.4%57.6%89.8%94.3%47.9%61.6%64.4%Reported by OpenAIOSWorld 2.083.7%82.9%54.5%64%Not yet published
GPT-5.3-CodexNot yet publishedNot yet publishedNot yet published91.5%Run by Artificial Analysis42.5%Not yet publishedNot yet published83.3%78.5%Not yet publishedNot yet published75.4%Run by Artificial AnalysisIFBench
Gemini 4 Argon57.6%61.8%86%96.4%57.1%77.5%69.2%Reported by GoogleOSWorld 2.079.7%Not yet published48.3%87%78.7%
Gemini 3.8 Flash19.7%56.6%68.4%95.4%47.8%59.9%59%Reported by GoogleOSWorld 2.081.3%85.6%50.5%86.1%Not yet published
Gemma 4 31BNot yet published45.5%Not yet published85.7%Run by Artificial AnalysisGPQA Diamond23.6%4.9%Not yet published69.7%73.4%10.4%Run by Epoch AISimpleQA Verified82.3%75.6%Run by Artificial AnalysisIFBench
Gemini 3.1 Pro Preview2.5%58.7%59.6%94.4%47%35.4%Not yet published82%82.4%54.9%59.9%71.4%
Gemini 3.6 Flash7.1%Run by Artificial AnalysisTerminal-Bench 4.053.4%58.9%94.1%40.8%53%83%Reported by GoogleOSWorld-Verified80%83.2%66.2%Run by Epoch AISimpleQA Verified87.2%Not yet published
DeepSeek V4.1 Flash31.8%55.9%71.2%93.2%43.3%66.6%69.3%Run by Moonshot AIOSWorld 2.182.7%80.6%39.6%Not yet published74.9%
DeepSeek V4 Pro14.1%51%64.6%91.7%41%56.7%Not yet published80.3%Not yet published45.2%72%Not yet published
Kimi K3.1 Pro63.6%60.8%Not yet published94.2%Reported by Moonshot AI52.5%73.3%79.6%Reported by Moonshot AIOSWorld 2.186.3%83.9%48.9%Not yet publishedNot yet published
Kimi K317.2%59.5%72.2%93.1%46.9%58.3%Not yet published88.7%80.5%50.6%Run by Epoch AISimpleQA Verified65.9%Not yet published
Kimi K2.7 Code1%Run by Artificial AnalysisTerminal-Bench 4.047.8%54%87.9%35%24.5%Not yet published79.3%Not yet published36.5%Run by Epoch AISimpleQA VerifiedNot yet published63.1%Run by Artificial AnalysisIFBench
Mistral Large 424.7%Run by Artificial AnalysisTerminal-Bench 4.0Not yet publishedNot yet publishedNot yet publishedNot yet published57.8%Not yet publishedNot yet publishedNot yet published26%Not yet publishedNot yet published
Mistral Medium 3.50%Run by Artificial AnalysisTerminal-Bench 4.040.2%Not yet published74.8%Run by Artificial AnalysisGPQA Diamond13.8%6.3%Not yet published69.3%64.9%23%70.5%68.8%Run by Artificial AnalysisIFBench
Grok 4.728.8%57.4%53%92.7%43.1%65.6%Not yet published76.7%Not yet published42.7%64.1%Not yet published
Grok 4.617.2%56.5%66%93.2%42.9%66.7%Not yet published80.3%Not yet published48.9%Run by Epoch AISimpleQA Verified83.6%Not yet published
Grok 4.58.6%55%57.2%93.4%42.7%57.9%Not yet published79.3%80.4%48.3%Run by Epoch AISimpleQA Verified93.1%Not yet published
MiniMax M3.1 Flash25.3%53.4%Not yet publishedNot yet published48%72.5%61.7%Reported by MiniMaxOSWorld 2.079.7%79.1%45.1%Not yet publishedNot yet published
Mercury 33%Not yet publishedNot yet publishedNot yet publishedNot yet publishedNot yet publishedNot yet publishedNot yet publishedNot yet publishedNot yet publishedNot yet publishedNot yet published
GLM-5.3-Flash37.4%51.6%55.8%90.2%39.9%60.4%59.1%Reported by Z.aiOSWorld 2.080%86%Run by Vals.aiMMMU-Pro28.2%65.4%Not yet published
GLM-5.338.9%59%68.8%90.9%42.3%62.2%Not yet published79.7%Not yet published32.3%71.5%Not yet published
GLM-5.21%Run by Artificial AnalysisTerminal-Bench 4.051.2%59.2%Not yet published41.1%28.4%Not yet published78.3%Not yet published34.2%Run by Epoch AISimpleQA VerifiedNot yet published73.3%Run by Artificial AnalysisIFBench
Muse Spark Open6.6%58.4%69.8%88.6%43.9%49.7%80.4%Reported by MetaOSWorld-Verified78.3%80.9%42%Not yet published74.2%
Muse Wonder49.5%60.3%82.1%Not yet published50.1%68.1%72.6%Reported by Meta80.7%Not yet published38%42.7%Not yet published
Muse Spark 1.310.6%58.8%74%93.5%Run by Artificial AnalysisGPQA Diamond48.7%57.9%69.6%83%82%38%25.8%Not yet published
Llama 4 ScoutNot yet published21.3%7.8%Run by Epoch AIOTIS Mock AIME 2024-202551.8%3.8%0.2%Not yet published27.7%52.9%Not yet published70.1%39.5%Run by Artificial AnalysisIFBench
Llama 4 MaverickNot yet published31.7%20.6%Run by Epoch AIOTIS Mock AIME 2024-202567%4.9%0.2%Not yet published50%62.1%Not yet published58.3%43%Run by Artificial AnalysisIFBench
MiMo-V2.6-Pro31.3%60.9%Not yet publishedNot yet published49.4%58.6%82%Reported by XiaomiOSWorld-Verified86.3%Not yet published34.8%66.5%Not yet published
MiMo-V2.6-Flash24.2%51.3%Not yet publishedNot yet published35.1%64.1%80.8%Reported by XiaomiOSWorld-Verified74.3%73.1%27.2%66.9%Not yet published
Ling 3.1 Flash32.8%Run by Artificial AnalysisTerminal-Bench 4.056.8%Not yet publishedNot yet published43.5%67.5%Not yet published82.3%Not yet published26.5%Not yet publishedNot yet published
Step 5 Preview31.8%58.9%Not yet published93.5%Reported by StepFunGPQA Diamond46.5%51%Not yet published88.3%76.4%36.4%59.7%Not yet published

Agentic codingTerminal-Bench 4.0

  1. Claude Opus 5.565.2%
  2. Claude Sonnet 5.564.1%
  3. Kimi K3.1 Pro63.6%
  4. GPT-6 Astra59.6%
  5. Claude Fable 5.158.1%
  6. Gemini 4 Argon57.6%
  7. GPT-6.1 Sol55.1%
  8. Claude Opus 553.5%
  9. GPT-6.1 Terra49.5%
  10. Muse Wonder49.5%
  11. GPT-6 Sol44.4%
  12. Claude Fable 541.4%
  13. Qwen3.8 Max38.9%
  14. GLM-5.338.9%
  15. GPT-5.6 Sol37.9%
  16. GLM-5.3-Flash37.4%
  17. Claude Haiku 5.536.4%
  18. DeepSeek V4.1 Flash31.8%
  19. Step 5 Preview31.8%
  20. MiMo-V2.6-Pro31.3%
  21. Grok 4.728.8%
  22. MiniMax M3.1 Flash25.3%
  23. MiMo-V2.6-Flash24.2%
  24. Claude Opus 4.823.2%
  25. GPT-5.6 Terra22.7%
  26. Gemini 3.8 Flash19.7%
  27. Kimi K317.2%
  28. Grok 4.617.2%
  29. DeepSeek V4 Pro14.1%
  30. GPT-6 Luna13.6%
  31. GPT-5.6 Luna11.6%
  32. Claude Sonnet 510.6%
  33. Muse Spark 1.310.6%
  34. Grok 4.58.6%
  35. Muse Spark Open6.6%
  36. Mercury 33%
  37. Gemini 3.1 Pro Preview2.5%
  38. Ling 3.1 Flash32.8%Artificial Analysis's run, Terminal-Bench 4.0
  39. Mistral Large 424.7%Artificial Analysis's run, Terminal-Bench 4.0
  40. GPT-5.514.6%Artificial Analysis's run, Terminal-Bench 4.0
  41. gpt-oss-120b14.2%Terminal-Bench's run, Terminal-Bench 2.0
  42. Gemini 3.6 Flash7.1%Artificial Analysis's run, Terminal-Bench 4.0
  43. Qwen3.8 27B5.6%Artificial Analysis's run, Terminal-Bench 4.0
  44. Claude Sonnet 4.63%Artificial Analysis's run, Terminal-Bench 4.0
  45. Kimi K2.7 Code1%Artificial Analysis's run, Terminal-Bench 4.0
  46. GLM-5.21%Artificial Analysis's run, Terminal-Bench 4.0
  47. Mistral Medium 3.50%Artificial Analysis's run, Terminal-Bench 4.0

The other models in the grid have no result on Terminal-Bench 4.0 yet.

CodeSciCode

  1. Claude Opus 5.566.9%
  2. Claude Fable 5.163.1%
  3. Gemini 4 Argon61.8%
  4. Claude Sonnet 5.561%
  5. Claude Fable 561%
  6. MiMo-V2.6-Pro60.9%
  7. Kimi K3.1 Pro60.8%
  8. Muse Wonder60.3%
  9. Kimi K359.5%
  10. GLM-5.359%
  11. Step 5 Preview58.9%
  12. Muse Spark 1.358.8%
  13. Gemini 3.1 Pro Preview58.7%
  14. Muse Spark Open58.4%
  15. GPT-6 Sol57.6%
  16. Grok 4.757.4%
  17. GPT-5.6 Sol57.1%
  18. Ling 3.1 Flash56.8%
  19. Gemini 3.8 Flash56.6%
  20. GPT-6 Astra56.5%
  21. Grok 4.656.5%
  22. Claude Opus 556.4%
  23. DeepSeek V4.1 Flash55.9%
  24. GPT-5.555.8%
  25. GPT-6.1 Terra55.1%
  26. Claude Haiku 5.555%
  27. GPT-5.6 Terra55%
  28. Grok 4.555%
  29. GPT-6 Luna54.6%
  30. Claude Opus 4.754.5%
  31. Claude Opus 4.854.4%
  32. Claude Sonnet 554.3%
  33. GPT-6.1 Sol54.2%
  34. GPT-5.6 Luna53.6%
  35. Gemini 3.6 Flash53.4%
  36. MiniMax M3.1 Flash53.4%
  37. Qwen3.8 Max52.1%
  38. GLM-5.3-Flash51.6%
  39. MiMo-V2.6-Flash51.3%
  40. GLM-5.251.2%
  41. DeepSeek V4 Pro51%
  42. Claude Sonnet 4.650.1%
  43. Kimi K2.7 Code47.8%
  44. Qwen3.8 27B46.6%
  45. Claude Sonnet 4.545.7%
  46. Gemma 4 31B45.5%
  47. Claude Haiku 4.542.2%
  48. Mistral Medium 3.540.2%
  49. gpt-oss-120b34%
  50. Llama 4 Maverick31.7%
  51. Llama 4 Scout21.3%

The other models in the grid have no result on SciCode yet.

MathFrontierMath Tiers 1-3

  1. GPT-6.1 Sol93.7%
  2. GPT-6 Astra93.7%
  3. Claude Opus 5.591.2%
  4. GPT-6.1 Terra90.9%
  5. Claude Fable 5.190.2%
  6. GPT-6 Sol89.8%
  7. GPT-5.6 Sol89.1%
  8. Claude Sonnet 5.588.8%
  9. Claude Fable 587%
  10. GPT-5.6 Terra86%
  11. Gemini 4 Argon86%
  12. Claude Opus 585.6%
  13. GPT-5.585.3%
  14. GPT-5.6 Luna82.1%
  15. Muse Wonder82.1%
  16. Claude Opus 4.880%
  17. GPT-6 Luna78.9%
  18. GPT-5.478.6%
  19. Claude Haiku 5.575.1%
  20. Muse Spark 1.374%
  21. Kimi K372.2%
  22. DeepSeek V4.1 Flash71.2%
  23. Claude Opus 4.770.2%
  24. Muse Spark Open69.8%
  25. GLM-5.368.8%
  26. Gemini 3.8 Flash68.4%
  27. Claude Opus 4.666%
  28. Grok 4.666%
  29. Qwen3.8 Max65.6%
  30. Claude Sonnet 565.6%
  31. DeepSeek V4 Pro64.6%
  32. Gemini 3.1 Pro Preview59.6%
  33. GLM-5.259.2%
  34. Gemini 3.6 Flash58.9%
  35. Grok 4.557.2%
  36. GLM-5.3-Flash55.8%
  37. Kimi K2.7 Code54%
  38. Grok 4.753%
  39. Claude Opus 4.534.4%
  40. Claude Sonnet 4.523.9%
  41. gpt-oss-120b88.9%Epoch AI's run, OTIS Mock AIME 2024-2025
  42. Claude Sonnet 4.671.1%Epoch AI's run, OTIS Mock AIME 2024-2025
  43. Claude Haiku 4.566.7%Epoch AI's run, OTIS Mock AIME 2024-2025
  44. Llama 4 Maverick20.6%Epoch AI's run, OTIS Mock AIME 2024-2025
  45. Llama 4 Scout7.8%Epoch AI's run, OTIS Mock AIME 2024-2025

The other models in the grid have no result on FrontierMath Tiers 1-3 yet.

ScienceGPQA Diamond

  1. Gemini 4 Argon96.4%
  2. GPT-6 Astra95.8%
  3. Claude Sonnet 5.595.6%
  4. GPT-6.1 Sol95.4%
  5. Gemini 3.8 Flash95.4%
  6. Gemini 3.1 Pro Preview94.4%
  7. GPT-6.1 Terra94.3%
  8. GPT-6 Sol94.3%
  9. Gemini 3.6 Flash94.1%
  10. Claude Opus 593.9%
  11. GPT-5.6 Sol93.5%
  12. Grok 4.593.4%
  13. GPT-5.6 Terra93.3%
  14. GPT-5.493.3%
  15. DeepSeek V4.1 Flash93.2%
  16. Grok 4.693.2%
  17. Kimi K393.1%
  18. Grok 4.792.7%
  19. Qwen3.8 Max92.3%
  20. DeepSeek V4 Pro91.7%
  21. GPT-5.6 Luna91.6%
  22. Claude Opus 4.891%
  23. GLM-5.390.9%
  24. Claude Opus 5.590.6%
  25. GPT-6 Luna90.5%
  26. GLM-5.3-Flash90.2%
  27. Claude Haiku 5.589.6%
  28. Muse Spark Open88.6%
  29. Claude Opus 4.688.4%
  30. Kimi K2.7 Code87.9%
  31. Claude Opus 4.786.4%
  32. Claude Opus 4.586%
  33. Claude Fable 585.9%
  34. Claude Sonnet 4.582.3%
  35. Claude Sonnet 580.3%
  36. Claude Sonnet 4.678.8%
  37. gpt-oss-120b75.8%
  38. Claude Haiku 4.571.2%
  39. Llama 4 Maverick67%
  40. Llama 4 Scout51.8%
  41. Kimi K3.1 Pro94.2%Moonshot AI's own figure
  42. GPT-5.593.5%Artificial Analysis's run
  43. Muse Spark 1.393.5%Artificial Analysis's run, GPQA Diamond
  44. Step 5 Preview93.5%StepFun's own figure, GPQA Diamond
  45. Claude Fable 5.193.4%Vals.ai's run
  46. GPT-5.3-Codex91.5%Artificial Analysis's run
  47. Qwen3.8 27B90.5%Artificial Analysis's run, GPQA Diamond
  48. Gemma 4 31B85.7%Artificial Analysis's run, GPQA Diamond
  49. Mistral Medium 3.574.8%Artificial Analysis's run, GPQA Diamond

The other models in the grid have no result on GPQA Diamond yet.

Hard reasoningHumanity's Last Exam

  1. Claude Opus 5.561.4%
  2. Claude Fable 5.159.1%
  3. Gemini 4 Argon57.1%
  4. Claude Fable 555.5%
  5. Claude Sonnet 5.555%
  6. Claude Opus 554.9%
  7. GPT-6 Astra54.7%
  8. GPT-6.1 Sol52.9%
  9. Kimi K3.1 Pro52.5%
  10. Muse Wonder50.1%
  11. GPT-5.6 Sol49.5%
  12. MiMo-V2.6-Pro49.4%
  13. Claude Opus 4.848.7%
  14. Muse Spark 1.348.7%
  15. MiniMax M3.1 Flash48%
  16. GPT-6 Sol47.9%
  17. Gemini 3.8 Flash47.8%
  18. Gemini 3.1 Pro Preview47%
  19. Kimi K346.9%
  20. Step 5 Preview46.5%
  21. GPT-6.1 Terra46.2%
  22. GPT-5.545.8%
  23. Claude Haiku 5.544.4%
  24. Muse Spark Open43.9%
  25. GPT-5.443.7%
  26. Ling 3.1 Flash43.5%
  27. DeepSeek V4.1 Flash43.3%
  28. Qwen3.8 Max43.1%
  29. Grok 4.743.1%
  30. GPT-5.6 Terra42.9%
  31. Grok 4.642.9%
  32. Grok 4.542.7%
  33. GPT-5.3-Codex42.5%
  34. Claude Opus 4.742.3%
  35. GLM-5.342.3%
  36. Claude Sonnet 541.3%
  37. GLM-5.241.1%
  38. DeepSeek V4 Pro41%
  39. Gemini 3.6 Flash40.8%
  40. Claude Opus 4.639.9%
  41. GLM-5.3-Flash39.9%
  42. GPT-5.6 Luna39.5%
  43. GPT-6 Luna38.5%
  44. MiMo-V2.6-Flash35.1%
  45. Kimi K2.7 Code35%
  46. Qwen3.8 27B33.9%
  47. Claude Sonnet 4.633.6%
  48. Claude Opus 4.530.1%
  49. Gemma 4 31B23.6%
  50. gpt-oss-120b19.6%
  51. Claude Sonnet 4.517.8%
  52. Mistral Medium 3.513.8%
  53. Claude Haiku 4.510.4%
  54. Llama 4 Maverick4.9%
  55. Llama 4 Scout3.8%

The other models in the grid have no result on Humanity's Last Exam yet.

Tool useAutomationBench-AA

  1. Gemini 4 Argon77.5%
  2. Kimi K3.1 Pro73.3%
  3. MiniMax M3.1 Flash72.5%
  4. Claude Sonnet 5.571.8%
  5. Claude Opus 5.569.5%
  6. GPT-6 Astra68.5%
  7. Muse Wonder68.1%
  8. Ling 3.1 Flash67.5%
  9. Grok 4.666.7%
  10. DeepSeek V4.1 Flash66.6%
  11. Grok 4.765.6%
  12. GPT-6.1 Sol64.9%
  13. MiMo-V2.6-Flash64.1%
  14. GLM-5.362.2%
  15. GPT-6 Sol61.6%
  16. GLM-5.3-Flash60.4%
  17. GPT-6.1 Terra60.2%
  18. GPT-5.6 Sol60.1%
  19. Gemini 3.8 Flash59.9%
  20. GPT-5.6 Terra59.6%
  21. Claude Fable 5.159.4%
  22. MiMo-V2.6-Pro58.6%
  23. Kimi K358.3%
  24. Grok 4.557.9%
  25. Muse Spark 1.357.9%
  26. Mistral Large 457.8%
  27. DeepSeek V4 Pro56.7%
  28. Claude Opus 556.6%
  29. Qwen3.8 Max56.2%
  30. Claude Fable 554.1%
  31. GPT-6 Luna53.2%
  32. Gemini 3.6 Flash53%
  33. Step 5 Preview51%
  34. GPT-5.6 Luna50.2%
  35. Muse Spark Open49.7%
  36. Qwen3.8 27B48.2%
  37. GPT-5.547.3%
  38. Claude Opus 4.845.6%
  39. Claude Sonnet 536.5%
  40. Claude Haiku 5.535.4%
  41. Gemini 3.1 Pro Preview35.4%
  42. GLM-5.228.4%
  43. Kimi K2.7 Code24.5%
  44. Claude Sonnet 4.620.1%
  45. Claude Sonnet 4.514.1%
  46. Mistral Medium 3.56.3%
  47. Gemma 4 31B4.9%
  48. Claude Haiku 4.53.2%
  49. gpt-oss-120b0.2%
  50. Llama 4 Scout0.2%
  51. Llama 4 Maverick0.2%

The other models in the grid have no result on AutomationBench-AA yet.

Computer useOSWorld 2.1

  1. Claude Opus 577.7%
  2. Muse Spark 1.369.6%
  3. Qwen3.8 27B84.3%Alibaba's own figure, OSWorld-Verified
  4. Gemini 3.6 Flash83%Google's own figure, OSWorld-Verified
  5. MiMo-V2.6-Pro82%Xiaomi's own figure, OSWorld-Verified
  6. Claude Opus 5.581.8%Anthropic's own figure
  7. MiMo-V2.6-Flash80.8%Xiaomi's own figure, OSWorld-Verified
  8. Claude Fable 5.180.7%Anthropic's own figure
  9. Muse Spark Open80.4%Meta's own figure, OSWorld-Verified
  10. Claude Sonnet 5.580.1%Anthropic's own figure
  11. Kimi K3.1 Pro79.6%Moonshot AI's own figure, OSWorld 2.1
  12. GPT-6.1 Sol78%OpenAI's own figure, OSWorld 2.1
  13. GPT-6 Astra75.8%Meta's run
  14. GPT-6.1 Terra74.2%OpenAI's own figure, OSWorld 2.1
  15. Muse Wonder72.6%Meta's own figure
  16. Claude Haiku 5.572.4%Anthropic's own figure, OSWorld 2.1
  17. DeepSeek V4.1 Flash69.3%Moonshot AI's run, OSWorld 2.1
  18. Gemini 4 Argon69.2%Google's own figure, OSWorld 2.0
  19. GPT-6 Sol64.4%OpenAI's own figure, OSWorld 2.0
  20. GPT-5.6 Sol62.7%OSWorld team's run, OSWorld 2.0
  21. MiniMax M3.1 Flash61.7%MiniMax's own figure, OSWorld 2.0
  22. GLM-5.3-Flash59.1%Z.ai's own figure, OSWorld 2.0
  23. Gemini 3.8 Flash59%Google's own figure, OSWorld 2.0
  24. Claude Sonnet 557%Anthropic's own figure
  25. Claude Opus 4.854.8%OSWorld team's run, OSWorld 2.0
  26. GPT-5.549.5%OSWorld team's run, OSWorld 2.0
  27. Claude Opus 4.748.9%OSWorld team's run, OSWorld 2.0
  28. GPT-6 Luna48.9%Anthropic's run, OSWorld 2.1
  29. Claude Sonnet 4.641.5%OSWorld team's run, OSWorld 2.0

The other models in the grid have no result on OSWorld 2.1 yet.

Long contextAA-LCR v1.1

  1. Kimi K388.7%
  2. Step 5 Preview88.3%
  3. Kimi K3.1 Pro86.3%
  4. MiMo-V2.6-Pro86.3%
  5. Claude Fable 5.185.3%
  6. Claude Opus 5.584.7%
  7. GPT-5.584.3%
  8. GPT-5.6 Sol84%
  9. GPT-5.6 Luna83.7%
  10. GPT-6 Sol83.7%
  11. GPT-6 Luna83.3%
  12. GPT-5.3-Codex83.3%
  13. GPT-6.1 Sol83%
  14. GPT-5.6 Terra83%
  15. Muse Spark 1.383%
  16. Claude Haiku 5.582.7%
  17. Claude Sonnet 5.582.7%
  18. GPT-6.1 Terra82.7%
  19. DeepSeek V4.1 Flash82.7%
  20. Claude Fable 582.3%
  21. Ling 3.1 Flash82.3%
  22. Qwen3.8 27B82%
  23. Claude Sonnet 582%
  24. GPT-5.482%
  25. Gemini 3.1 Pro Preview82%
  26. Gemini 3.8 Flash81.3%
  27. GPT-6 Astra80.7%
  28. Muse Wonder80.7%
  29. Qwen3.8 Max80.3%
  30. DeepSeek V4 Pro80.3%
  31. Grok 4.680.3%
  32. Claude Sonnet 4.680%
  33. Gemini 3.6 Flash80%
  34. GLM-5.3-Flash80%
  35. Gemini 4 Argon79.7%
  36. MiniMax M3.1 Flash79.7%
  37. GLM-5.379.7%
  38. Claude Opus 579.3%
  39. Kimi K2.7 Code79.3%
  40. Grok 4.579.3%
  41. Claude Opus 4.778.7%
  42. GLM-5.278.3%
  43. Muse Spark Open78.3%
  44. Claude Opus 4.678%
  45. Claude Opus 4.877.7%
  46. Claude Opus 4.577.3%
  47. Grok 4.776.7%
  48. Claude Haiku 4.574.3%
  49. MiMo-V2.6-Flash74.3%
  50. Claude Sonnet 4.572.3%
  51. Gemma 4 31B69.7%
  52. Mistral Medium 3.569.3%
  53. gpt-oss-120b52%
  54. Llama 4 Maverick50%
  55. Llama 4 Scout27.7%

The other models in the grid have no result on AA-LCR v1.1 yet.

MultimodalMMMU-Pro

  1. Claude Opus 5.587.7%
  2. GPT-6 Astra86.9%
  3. GPT-6.1 Sol86%
  4. Gemini 3.8 Flash85.6%
  5. Claude Opus 584.7%
  6. GPT-6.1 Terra84.7%
  7. Kimi K3.1 Pro83.9%
  8. GPT-5.6 Sol83.4%
  9. Gemini 3.6 Flash83.2%
  10. GPT-6 Sol82.9%
  11. Qwen3.8 Max82.8%
  12. Gemini 3.1 Pro Preview82.4%
  13. Muse Spark 1.382%
  14. Muse Spark Open80.9%
  15. GPT-5.6 Terra80.7%
  16. DeepSeek V4.1 Flash80.6%
  17. Kimi K380.5%
  18. Grok 4.580.4%
  19. GPT-5.579.9%
  20. GPT-6 Luna79.7%
  21. MiniMax M3.1 Flash79.1%
  22. Claude Opus 4.778.8%
  23. GPT-5.6 Luna78.6%
  24. GPT-5.3-Codex78.5%
  25. GPT-5.478.4%
  26. Claude Sonnet 577.3%
  27. Step 5 Preview76.4%
  28. Qwen3.8 27B76.3%
  29. Claude Opus 4.675.4%
  30. Claude Opus 4.574%
  31. Gemma 4 31B73.4%
  32. Claude Sonnet 4.673.3%
  33. MiMo-V2.6-Flash73.1%
  34. Claude Sonnet 4.568.7%
  35. Mistral Medium 3.564.9%
  36. Llama 4 Maverick62.1%
  37. Claude Haiku 4.558.6%
  38. Llama 4 Scout52.9%
  39. Claude Fable 5.186.3%Moonshot AI's run
  40. GLM-5.3-Flash86%Vals.ai's run, MMMU-Pro

The other models in the grid have no result on MMMU-Pro yet.

FactualityAA-Omniscience Accuracy

  1. Claude Fable 5.168.4%
  2. Claude Opus 5.565.5%
  3. Claude Fable 565.4%
  4. GPT-6.1 Sol61.8%
  5. GPT-6 Astra61.6%
  6. Claude Opus 560.9%
  7. Gemini 3.1 Pro Preview54.9%
  8. GPT-6 Sol54.5%
  9. Claude Sonnet 5.551.5%
  10. Gemini 3.8 Flash50.5%
  11. Claude Opus 4.748.9%
  12. Kimi K3.1 Pro48.9%
  13. Claude Opus 4.848.8%
  14. Gemini 4 Argon48.3%
  15. GPT-6.1 Terra47.6%
  16. Claude Opus 4.647%
  17. Claude Opus 4.546.6%
  18. DeepSeek V4 Pro45.2%
  19. MiniMax M3.1 Flash45.1%
  20. Grok 4.742.7%
  21. GPT-5.6 Terra42.3%
  22. Muse Spark Open42%
  23. Claude Sonnet 4.640.9%
  24. Claude Sonnet 540.1%
  25. GPT-6 Luna39.6%
  26. DeepSeek V4.1 Flash39.6%
  27. Muse Wonder38%
  28. Muse Spark 1.338%
  29. Step 5 Preview36.4%
  30. MiMo-V2.6-Pro34.8%
  31. Claude Haiku 5.534.7%
  32. Claude Sonnet 4.532.9%
  33. GLM-5.332.3%
  34. Qwen3.8 Max29.5%
  35. GLM-5.3-Flash28.2%
  36. MiMo-V2.6-Flash27.2%
  37. Ling 3.1 Flash26.5%
  38. Mistral Large 426%
  39. Mistral Medium 3.523%
  40. Claude Haiku 4.518%
  41. Qwen3.8 27B17.3%
  42. GPT-5.6 Sol69.7%Epoch AI's run, SimpleQA Verified
  43. Gemini 3.6 Flash66.2%Epoch AI's run, SimpleQA Verified
  44. GPT-5.563%Epoch AI's run, SimpleQA Verified
  45. Kimi K350.6%Epoch AI's run, SimpleQA Verified
  46. Grok 4.648.9%Epoch AI's run, SimpleQA Verified
  47. Grok 4.548.3%Epoch AI's run, SimpleQA Verified
  48. GPT-5.6 Luna41%Epoch AI's run, SimpleQA Verified
  49. Kimi K2.7 Code36.5%Epoch AI's run, SimpleQA Verified
  50. GLM-5.234.2%Epoch AI's run, SimpleQA Verified
  51. gpt-oss-120b16.8%OpenAI's own figure, SimpleQA
  52. Gemma 4 31B10.4%Epoch AI's run, SimpleQA Verified

The other models in the grid have no result on AA-Omniscience Accuracy yet.

Answers in fullSpeechMap

  1. Grok 4.593.1%
  2. Gemini 3.6 Flash87.2%
  3. Gemini 4 Argon87%
  4. Gemini 3.8 Flash86.1%
  5. Grok 4.683.6%
  6. Claude Opus 5.582.8%
  7. Gemma 4 31B82.3%
  8. Claude Sonnet 5.581.6%
  9. Claude Haiku 5.578.8%
  10. Claude Opus 573.7%
  11. Claude Fable 5.172.4%
  12. DeepSeek V4 Pro72%
  13. GLM-5.371.5%
  14. Mistral Medium 3.570.5%
  15. Claude Opus 4.770.2%
  16. Llama 4 Scout70.1%
  17. GPT-6.1 Sol69.5%
  18. GPT-6 Astra68.2%
  19. Claude Opus 4.867.4%
  20. MiMo-V2.6-Flash66.9%
  21. MiMo-V2.6-Pro66.5%
  22. Kimi K365.9%
  23. GLM-5.3-Flash65.4%
  24. Qwen3.8 Max65.3%
  25. Claude Fable 564.8%
  26. Grok 4.764.1%
  27. GPT-6 Sol64%
  28. GPT-6 Luna63.4%
  29. Claude Opus 4.659.9%
  30. Gemini 3.1 Pro Preview59.9%
  31. Step 5 Preview59.7%
  32. Llama 4 Maverick58.3%
  33. Claude Opus 4.557.5%
  34. Muse Wonder42.7%
  35. gpt-oss-120b36.3%
  36. Claude Sonnet 4.635.8%
  37. Claude Sonnet 4.531.4%
  38. Muse Spark 1.325.8%
  39. Claude Haiku 4.524.7%

The other models in the grid have no result on SpeechMap yet.

Instruction followingMultiChallenge

  1. GPT-6.1 Sol82%
  2. Gemini 4 Argon78.7%
  3. DeepSeek V4.1 Flash74.9%
  4. GPT-6.1 Terra74.2%
  5. Muse Spark Open74.2%
  6. Gemini 3.1 Pro Preview71.4%
  7. Claude Haiku 5.567.7%
  8. Claude Opus 4.559%
  9. Claude Sonnet 4.555.3%
  10. Claude Haiku 4.550.5%
  11. gpt-oss-120b45.3%
  12. Claude Opus 4.637.2%
  13. Qwen3.8 27B79.5%Alibaba's own figure, IFBench
  14. GPT-5.575.9%Artificial Analysis's run, IFBench
  15. Gemma 4 31B75.6%Artificial Analysis's run, IFBench
  16. GPT-5.3-Codex75.4%Artificial Analysis's run, IFBench
  17. GPT-5.473.9%Artificial Analysis's run, IFBench
  18. GLM-5.273.3%Artificial Analysis's run, IFBench
  19. GPT-5.6 Sol72.7%Artificial Analysis's run, IFBench
  20. GPT-5.6 Terra71.2%Artificial Analysis's run, IFBench
  21. Mistral Medium 3.568.8%Artificial Analysis's run, IFBench
  22. Claude Fable 563.5%Artificial Analysis's run, IFBench
  23. Kimi K2.7 Code63.1%Artificial Analysis's run, IFBench
  24. Claude Opus 4.862.2%Artificial Analysis's run, IFBench
  25. Claude Opus 4.758.6%Artificial Analysis's run, IFBench
  26. Claude Sonnet 4.656.6%Artificial Analysis's run, IFBench
  27. Llama 4 Maverick43%Artificial Analysis's run, IFBench
  28. Llama 4 Scout39.5%Artificial Analysis's run, IFBench

The other models in the grid have no result on MultiChallenge yet.

Releases

Every Qwen release we track, newest first.

  1. Qwen3.8 27BAlibaba's dense 27B Qwen3.8 vision-language model for coding and agentic work, with a 262,144-token native context. Weights under the Apache 2.0 license; on Alibaba Cloud Model Studio as qwen3.8-27b.CurrentAug 14, 2026
  2. Qwen3.8 MaxAlibaba's flagship Qwen3.8 model on Model Studio, with text, image and video input, on Alibaba Cloud Model Studio as qwen3.8-max. Since September 5, 2026 that endpoint serves the qwen3.8-max-0902 snapshot of September 2, at the same price.CurrentAug 2, 2026

What practitioners report

Strongest on Long context, weakest on Developer experience.

Each figure is the share of accounts naming Qwen3.8 27B that speak well of it on that axis. An axis is scored once praise and concerns are both on record from at least five accounts by three authors. An account from the maker, a company that sells or serves the model, or the maker's expert program is listed and counts toward no figure.

Share of practitioner accounts favorable to Qwen3.8 27B on each of seven axes, with the accounts behind each figure
AxisFavorableAccounts
Quality and reasoningWhether its answers and code are right on the author's own task: reasoning, accuracy, instruction following, and consistency.71%24 favorable, 10 not, 34 authors
Tool use and structured outputWhether it calls tools correctly, returns valid structured output, and keeps agent loops on track.46%6 favorable, 7 not, 13 authors
Long contextHow it holds up over long documents, large codebases, long sessions, and many images or PDFs.75%6 favorable, 2 not, 8 authors
Developer experienceHow it is to build with: prompting, steering, effort settings, and support across SDKs and providers.17%1 favorable, 5 not, 6 authors
CostWhat the work costs: price per task, token use, and what caching and batch save.Needs 5 accounts from 3 authors2 favorable, 2 not, 4 authors
SpeedResponse time, time to first token, and throughput.33%7 favorable, 14 not, 21 authors
ReliabilityWhether requests keep succeeding: overloaded errors, outages, rate limits, and changes that broke what worked.Only concerns on record0 favorable, 9 not, 9 authors
Every model on these seven axes
Share of accounts favorable on each axis for Qwen3.8 27B and every other model GILD has accounts for
AxisQwen3.8 27BQwen3.8 MaxClaude Haiku 5.5Claude Sonnet 5.5Claude Opus 5.5Claude Fable 5.1Claude Opus 5Claude Sonnet 5Claude Fable 5Claude Opus 4.8Claude Opus 4.7Claude Sonnet 4.6Claude Opus 4.6Claude Opus 4.5Claude Haiku 4.5Claude Sonnet 4.5GPT-6.1 TerraGPT-6.1 SolGPT-6 LunaGPT-6 AstraGPT-5.6 SolGPT-5.6 TerraGPT-5.6 LunaGPT-5.5GPT-5.4gpt-oss-120bGPT-6 SolGPT-5.3-CodexGemini 4 ArgonGemini 3.8 FlashGemma 4 31BGemini 3.1 Pro PreviewGemini 3.6 FlashDeepSeek V4.1 FlashDeepSeek V4 ProKimi K3.1 ProKimi K3Kimi K2.7 CodeMistral Large 4Mistral Medium 3.5Grok 4.7Grok 4.6Grok 4.5MiniMax M3.1 FlashMercury 3GLM-5.3-FlashGLM-5.3GLM-5.2Muse Spark OpenMuse WonderMuse Spark 1.3Llama 4 ScoutLlama 4 MaverickMiMo-V2.6-ProMiMo-V2.6-FlashLing 3.1 FlashStep 5 Preview
Quality and reasoning71%24 favorable, 10 not, 34 authors57%4 favorable, 3 not, 7 authors73%19 favorable, 7 not, 26 authors64%14 favorable, 8 not, 22 authors68%17 favorable, 8 not, 25 authors54%15 favorable, 13 not, 28 authors50%6 favorable, 6 not, 12 authors56%9 favorable, 7 not, 16 authors68%13 favorable, 6 not, 18 authors68%17 favorable, 8 not, 25 authors57%12 favorable, 9 not, 21 authors67%12 favorable, 6 not, 18 authors70%19 favorable, 8 not, 27 authors62%13 favorable, 8 not, 20 authors58%14 favorable, 10 not, 24 authors69%11 favorable, 5 not, 16 authors33%3 favorable, 6 not, 8 authors55%12 favorable, 10 not, 22 authors55%16 favorable, 13 not, 29 authors44%12 favorable, 15 not, 26 authors33%3 favorable, 6 not, 9 authors38%3 favorable, 5 not, 8 authorsNeeds 5 accounts from 3 authors3 favorable, 1 not, 4 authors33%4 favorable, 8 not, 12 authors33%2 favorable, 4 not, 6 authors58%22 favorable, 16 not, 38 authors31%10 favorable, 22 not, 32 authorsNeeds 5 accounts from 3 authors1 favorable, 3 not, 4 authors63%22 favorable, 13 not, 35 authors56%27 favorable, 21 not, 45 authors80%24 favorable, 6 not, 28 authors56%20 favorable, 16 not, 36 authorsNeeds 5 accounts from 3 authors1 favorable, 3 not, 4 authors61%25 favorable, 16 not, 40 authorsOnly concerns on record0 favorable, 4 not, 4 authors59%20 favorable, 14 not, 32 authorsOnly praise on record3 favorable, 0 not, 3 authors50%3 favorable, 3 not, 6 authors58%11 favorable, 8 not, 18 authors36%5 favorable, 9 not, 14 authors43%16 favorable, 21 not, 36 authors86%6 favorable, 1 not, 7 authors60%3 favorable, 2 not, 5 authors62%13 favorable, 8 not, 21 authors86%18 favorable, 3 not, 21 authors67%14 favorable, 7 not, 21 authors76%47 favorable, 15 not, 61 authors71%5 favorable, 2 not, 7 authorsNeeds 5 accounts from 3 authors2 favorable, 1 not, 3 authors20%1 favorable, 4 not, 5 authorsOnly praise on record2 favorable, 0 not, 2 authors53%8 favorable, 7 not, 15 authors44%8 favorable, 10 not, 17 authors22%2 favorable, 7 not, 9 authors40%4 favorable, 6 not, 10 authorsOnly concerns on record0 favorable, 1 not, 1 authorNeeds 5 accounts from 3 authors2 favorable, 1 not, 3 authors
Tool use and structured output46%6 favorable, 7 not, 13 authorsOnly praise on record1 favorable, 0 not, 1 author67%6 favorable, 3 not, 9 authors40%4 favorable, 6 not, 10 authors46%6 favorable, 7 not, 13 authors30%3 favorable, 7 not, 10 authorsNeeds 5 accounts from 3 authors3 favorable, 1 not, 4 authors70%7 favorable, 3 not, 10 authors57%4 favorable, 3 not, 7 authors50%4 favorable, 4 not, 8 authors50%4 favorable, 4 not, 8 authors29%2 favorable, 5 not, 7 authors33%4 favorable, 8 not, 12 authors70%7 favorable, 3 not, 10 authors43%3 favorable, 4 not, 6 authors50%3 favorable, 3 not, 6 authors40%2 favorable, 3 not, 5 authors40%4 favorable, 6 not, 10 authors33%3 favorable, 6 not, 9 authors62%8 favorable, 5 not, 13 authorsNo accounts yetOnly concerns on record0 favorable, 2 not, 2 authorsOnly praise on record1 favorable, 0 not, 1 authorOnly praise on record1 favorable, 0 not, 1 authorNo accounts yet27%4 favorable, 11 not, 15 authors56%5 favorable, 4 not, 9 authorsNo accounts yet25%2 favorable, 6 not, 8 authors47%7 favorable, 8 not, 15 authors67%4 favorable, 2 not, 6 authors54%7 favorable, 6 not, 12 authorsOnly concerns on record0 favorable, 1 not, 1 author59%10 favorable, 7 not, 17 authorsOnly concerns on record0 favorable, 1 not, 1 authorNeeds 5 accounts from 3 authors3 favorable, 1 not, 4 authorsOnly concerns on record0 favorable, 1 not, 1 authorNo accounts yet38%5 favorable, 8 not, 13 authorsOnly concerns on record0 favorable, 5 not, 5 authors38%6 favorable, 10 not, 16 authorsOnly praise on record2 favorable, 0 not, 2 authorsOnly concerns on record0 favorable, 1 not, 1 author67%4 favorable, 2 not, 6 authors80%8 favorable, 2 not, 10 authors33%2 favorable, 4 not, 6 authors57%4 favorable, 3 not, 7 authorsNo accounts yetNo accounts yetOnly concerns on record0 favorable, 1 not, 1 authorNeeds 5 accounts from 3 authors1 favorable, 1 not, 2 authorsNeeds 5 accounts from 3 authors1 favorable, 2 not, 3 authorsOnly concerns on record0 favorable, 4 not, 4 authorsOnly concerns on record0 favorable, 5 not, 5 authorsOnly concerns on record0 favorable, 1 not, 1 authorOnly concerns on record0 favorable, 1 not, 1 authorOnly concerns on record0 favorable, 1 not, 1 author
Long context75%6 favorable, 2 not, 8 authorsOnly praise on record1 favorable, 0 not, 1 author29%2 favorable, 5 not, 7 authors60%3 favorable, 2 not, 5 authors80%8 favorable, 2 not, 10 authors63%5 favorable, 3 not, 8 authors67%4 favorable, 2 not, 6 authors50%3 favorable, 3 not, 6 authors33%2 favorable, 4 not, 6 authors60%6 favorable, 4 not, 10 authors50%4 favorable, 4 not, 7 authors71%5 favorable, 2 not, 7 authors64%7 favorable, 4 not, 11 authors40%4 favorable, 6 not, 10 authors40%2 favorable, 3 not, 5 authors50%4 favorable, 4 not, 8 authorsOnly concerns on record0 favorable, 1 not, 1 author38%3 favorable, 5 not, 8 authors20%1 favorable, 4 not, 5 authors27%3 favorable, 8 not, 11 authorsNo accounts yetNo accounts yetOnly concerns on record0 favorable, 1 not, 1 authorOnly concerns on record0 favorable, 1 not, 1 authorNeeds 5 accounts from 3 authors2 favorable, 1 not, 3 authors25%2 favorable, 6 not, 8 authorsOnly concerns on record0 favorable, 3 not, 3 authorsNo accounts yet25%2 favorable, 6 not, 8 authors38%5 favorable, 8 not, 13 authors20%1 favorable, 4 not, 5 authors44%7 favorable, 9 not, 16 authorsOnly praise on record1 favorable, 0 not, 1 author58%7 favorable, 5 not, 12 authorsNo accounts yetNeeds 5 accounts from 3 authors3 favorable, 1 not, 4 authorsOnly praise on record1 favorable, 0 not, 1 authorOnly concerns on record0 favorable, 1 not, 1 authorNeeds 5 accounts from 3 authors1 favorable, 3 not, 4 authorsNo accounts yet42%5 favorable, 7 not, 12 authorsNo accounts yetNo accounts yet40%2 favorable, 3 not, 5 authors60%3 favorable, 2 not, 5 authorsNo accounts yet40%2 favorable, 3 not, 5 authorsNo accounts yetNo accounts yetOnly concerns on record0 favorable, 1 not, 1 authorOnly concerns on record0 favorable, 1 not, 1 authorOnly concerns on record0 favorable, 3 not, 3 authorsNeeds 5 accounts from 3 authors2 favorable, 1 not, 3 authorsNeeds 5 accounts from 3 authors1 favorable, 1 not, 2 authorsNeeds 5 accounts from 3 authors1 favorable, 1 not, 2 authorsNo accounts yetNo accounts yet
Developer experience17%1 favorable, 5 not, 6 authorsOnly concerns on record0 favorable, 1 not, 1 author25%2 favorable, 6 not, 8 authors38%5 favorable, 8 not, 13 authors38%6 favorable, 10 not, 16 authors50%7 favorable, 7 not, 14 authors25%3 favorable, 9 not, 12 authors29%2 favorable, 5 not, 7 authors25%3 favorable, 9 not, 12 authors53%10 favorable, 9 not, 19 authors29%4 favorable, 10 not, 14 authors57%4 favorable, 3 not, 7 authors45%5 favorable, 6 not, 11 authors67%8 favorable, 4 not, 12 authorsNeeds 5 accounts from 3 authors2 favorable, 2 not, 4 authors50%3 favorable, 3 not, 6 authorsOnly concerns on record0 favorable, 2 not, 2 authors27%3 favorable, 8 not, 11 authors25%2 favorable, 6 not, 8 authors41%7 favorable, 10 not, 17 authorsNo accounts yetOnly concerns on record0 favorable, 1 not, 1 authorNo accounts yetNo accounts yetNo accounts yet14%1 favorable, 6 not, 7 authors25%3 favorable, 9 not, 12 authorsNo accounts yet33%3 favorable, 6 not, 9 authors39%7 favorable, 11 not, 17 authorsOnly concerns on record0 favorable, 7 not, 7 authors40%8 favorable, 12 not, 20 authorsNo accounts yet35%8 favorable, 15 not, 23 authorsNo accounts yet9%1 favorable, 10 not, 11 authorsNo accounts yetOnly concerns on record0 favorable, 1 not, 1 author13%2 favorable, 14 not, 16 authorsOnly concerns on record0 favorable, 9 not, 9 authors25%5 favorable, 15 not, 20 authorsNo accounts yetNo accounts yetOnly concerns on record0 favorable, 7 not, 7 authors11%1 favorable, 8 not, 9 authorsOnly concerns on record0 favorable, 3 not, 3 authors33%5 favorable, 10 not, 15 authorsNo accounts yetOnly concerns on record0 favorable, 1 not, 1 authorOnly concerns on record0 favorable, 1 not, 1 authorNo accounts yetOnly concerns on record0 favorable, 6 not, 6 authorsOnly concerns on record0 favorable, 1 not, 1 authorOnly concerns on record0 favorable, 1 not, 1 authorOnly concerns on record0 favorable, 1 not, 1 authorNo accounts yetNo accounts yet
CostNeeds 5 accounts from 3 authors2 favorable, 2 not, 4 authorsOnly concerns on record0 favorable, 2 not, 2 authors69%20 favorable, 9 not, 29 authors45%10 favorable, 12 not, 22 authors56%15 favorable, 12 not, 27 authors20%4 favorable, 16 not, 20 authors43%3 favorable, 4 not, 7 authors39%7 favorable, 11 not, 17 authors8%1 favorable, 11 not, 12 authors40%6 favorable, 9 not, 14 authors50%7 favorable, 7 not, 13 authors50%6 favorable, 6 not, 12 authors33%7 favorable, 14 not, 20 authors53%8 favorable, 7 not, 15 authors29%4 favorable, 10 not, 12 authors38%3 favorable, 5 not, 8 authors80%8 favorable, 2 not, 9 authors74%14 favorable, 5 not, 19 authors79%22 favorable, 6 not, 26 authors29%7 favorable, 17 not, 23 authors57%4 favorable, 3 not, 7 authorsOnly concerns on record0 favorable, 1 not, 1 author50%3 favorable, 3 not, 6 authorsNeeds 5 accounts from 3 authors1 favorable, 3 not, 4 authors63%5 favorable, 3 not, 8 authors60%3 favorable, 2 not, 5 authors55%12 favorable, 10 not, 21 authorsNeeds 5 accounts from 3 authors2 favorable, 1 not, 3 authors35%7 favorable, 13 not, 20 authors58%14 favorable, 10 not, 23 authorsOnly praise on record7 favorable, 0 not, 7 authors50%8 favorable, 8 not, 16 authorsOnly concerns on record0 favorable, 1 not, 1 author66%25 favorable, 13 not, 37 authorsOnly praise on record2 favorable, 0 not, 2 authors40%10 favorable, 15 not, 25 authors71%5 favorable, 2 not, 7 authorsNeeds 5 accounts from 3 authors1 favorable, 3 not, 4 authors27%3 favorable, 8 not, 11 authorsOnly concerns on record0 favorable, 8 not, 8 authors19%5 favorable, 21 not, 26 authors20%1 favorable, 4 not, 5 authorsOnly praise on record4 favorable, 0 not, 4 authors80%12 favorable, 3 not, 15 authors50%6 favorable, 6 not, 12 authors76%16 favorable, 5 not, 21 authors35%8 favorable, 15 not, 23 authors71%5 favorable, 2 not, 7 authorsOnly concerns on record0 favorable, 1 not, 1 authorOnly concerns on record0 favorable, 9 not, 9 authorsNeeds 5 accounts from 3 authors2 favorable, 1 not, 3 authorsOnly praise on record1 favorable, 0 not, 1 authorOnly praise on record1 favorable, 0 not, 1 authorNeeds 5 accounts from 3 authors1 favorable, 1 not, 2 authorsOnly praise on record3 favorable, 0 not, 3 authorsOnly praise on record1 favorable, 0 not, 1 authorOnly praise on record1 favorable, 0 not, 1 author
Speed33%7 favorable, 14 not, 21 authorsNeeds 5 accounts from 3 authors1 favorable, 3 not, 4 authors74%14 favorable, 5 not, 19 authors88%14 favorable, 2 not, 16 authors46%6 favorable, 7 not, 13 authors44%4 favorable, 5 not, 9 authors60%3 favorable, 2 not, 5 authors44%4 favorable, 5 not, 9 authors57%4 favorable, 3 not, 7 authors31%4 favorable, 9 not, 13 authors63%5 favorable, 3 not, 8 authors17%1 favorable, 5 not, 6 authors36%4 favorable, 7 not, 11 authors73%8 favorable, 3 not, 11 authors78%7 favorable, 2 not, 9 authors75%6 favorable, 2 not, 8 authors83%5 favorable, 1 not, 5 authors37%7 favorable, 12 not, 19 authors45%5 favorable, 6 not, 11 authors41%7 favorable, 10 not, 17 authorsOnly praise on record1 favorable, 0 not, 1 authorNeeds 5 accounts from 3 authors1 favorable, 1 not, 2 authors80%4 favorable, 1 not, 5 authorsNeeds 5 accounts from 3 authors1 favorable, 1 not, 2 authorsNo accounts yet85%23 favorable, 4 not, 27 authors57%4 favorable, 3 not, 7 authorsOnly praise on record1 favorable, 0 not, 1 author45%5 favorable, 6 not, 11 authors61%19 favorable, 12 not, 31 authors53%10 favorable, 9 not, 17 authors35%6 favorable, 11 not, 16 authorsNeeds 5 accounts from 3 authors2 favorable, 1 not, 3 authors68%21 favorable, 10 not, 30 authorsOnly concerns on record0 favorable, 1 not, 1 author21%4 favorable, 15 not, 19 authorsOnly concerns on record0 favorable, 2 not, 2 authorsNo accounts yet21%3 favorable, 11 not, 14 authorsOnly concerns on record0 favorable, 2 not, 2 authors29%5 favorable, 12 not, 17 authorsNeeds 5 accounts from 3 authors1 favorable, 2 not, 3 authorsOnly praise on record2 favorable, 0 not, 2 authors44%4 favorable, 5 not, 9 authors65%13 favorable, 7 not, 20 authors33%3 favorable, 6 not, 9 authors20%3 favorable, 12 not, 15 authorsNeeds 5 accounts from 3 authors1 favorable, 1 not, 2 authorsOnly concerns on record0 favorable, 1 not, 1 authorOnly concerns on record0 favorable, 1 not, 1 authorNeeds 5 accounts from 3 authors1 favorable, 1 not, 2 authorsNeeds 5 accounts from 3 authors3 favorable, 1 not, 4 authorsNo accounts yetOnly concerns on record0 favorable, 4 not, 4 authorsOnly concerns on record0 favorable, 1 not, 1 authorOnly concerns on record0 favorable, 2 not, 2 authorsOnly concerns on record0 favorable, 1 not, 1 author
ReliabilityOnly concerns on record0 favorable, 9 not, 9 authorsOnly concerns on record0 favorable, 1 not, 1 author20%1 favorable, 4 not, 5 authors18%2 favorable, 9 not, 11 authors13%1 favorable, 7 not, 8 authors9%1 favorable, 10 not, 11 authors13%1 favorable, 7 not, 8 authors63%5 favorable, 3 not, 7 authors22%2 favorable, 7 not, 9 authors17%2 favorable, 10 not, 12 authors18%2 favorable, 9 not, 11 authors17%1 favorable, 5 not, 6 authors20%3 favorable, 12 not, 15 authors38%6 favorable, 10 not, 16 authors29%2 favorable, 5 not, 7 authors40%4 favorable, 6 not, 10 authorsOnly concerns on record0 favorable, 1 not, 1 author13%1 favorable, 7 not, 8 authorsOnly concerns on record0 favorable, 5 not, 5 authors14%2 favorable, 12 not, 14 authorsOnly concerns on record0 favorable, 4 not, 4 authorsOnly concerns on record0 favorable, 1 not, 1 authorNo accounts yetOnly concerns on record0 favorable, 2 not, 2 authorsOnly concerns on record0 favorable, 4 not, 4 authorsOnly concerns on record0 favorable, 4 not, 4 authors17%1 favorable, 5 not, 6 authorsOnly concerns on record0 favorable, 3 not, 3 authors5%1 favorable, 18 not, 19 authors15%3 favorable, 17 not, 20 authorsOnly concerns on record0 favorable, 5 not, 5 authors14%3 favorable, 19 not, 22 authorsOnly concerns on record0 favorable, 1 not, 1 author42%5 favorable, 7 not, 12 authorsNeeds 5 accounts from 3 authors1 favorable, 1 not, 2 authors8%1 favorable, 12 not, 13 authorsOnly concerns on record0 favorable, 1 not, 1 authorNeeds 5 accounts from 3 authors1 favorable, 1 not, 2 authors38%3 favorable, 5 not, 8 authorsOnly concerns on record0 favorable, 4 not, 4 authors22%2 favorable, 7 not, 9 authorsOnly concerns on record0 favorable, 1 not, 1 authorNo accounts yet60%3 favorable, 2 not, 5 authors33%2 favorable, 4 not, 5 authorsOnly concerns on record0 favorable, 7 not, 7 authorsOnly concerns on record0 favorable, 4 not, 4 authorsOnly concerns on record0 favorable, 1 not, 1 authorNo accounts yetOnly concerns on record0 favorable, 2 not, 2 authorsOnly concerns on record0 favorable, 1 not, 1 authorNeeds 5 accounts from 3 authors1 favorable, 3 not, 4 authorsOnly concerns on record0 favorable, 1 not, 1 authorOnly concerns on record0 favorable, 2 not, 2 authorsNo accounts yetNo accounts yetOnly concerns on record0 favorable, 1 not, 1 author
The 72 accounts behind Qwen3.8 27B's figures

GILD searched Hacker News and GitHub for accounts that name Qwen3.8 27B.

Quality and reasoning 71%

  • Favorable
    This is by far the best pelican SVG I've been able to generate with a model that runs on a local machine

    Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things. Simon Willison, independent developer and blogger, self-described, Aug 16, 2026. Qwen3.8 27B as the 17GB Q4_K_M build in LM Studio on an M5 Max MacBook Pro and a DGX Spark.

  • Favorable
    Not only does my measurement of BF16 replicate the stated result, but, to my surprise, Q4_K_M does as well.

    Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses. Piotr Migdał, author on the Quesma blog, self-described, Aug 26, 2026. Qwen3.8 27B in BF16 and Unsloth Q8, Q4, Q2 and 1-bit quantizations on llama.cpp, GPQA Diamond, IFBench and Terminal-Bench 2.1.

  • Favorable
    the quality is much higher than Qwen 3.6 27B

    Qwen 3.8 27B. lelandbatey, Hacker News commenter, self-described, Aug 14, 2026. Qwen3.8 27B as the Unsloth 2-bit quant on a 5070ti with 16GB of VRAM at 128k context through llama.cpp.

  • Unfavorable
    it failed my basic prompt to compose a vim regex to match CamelCaseWords

    Qwen 3.8 27B. rasengan0, Hacker News commenter, self-described, Aug 15, 2026. Qwen3.8 27B as the Unsloth UD-Q4_K_XL and Ollama Q4_K_M quants on a Framework 12 i5 laptop with 48GB RAM.

  • Favorable
    I don't want to say it's as good as Claude (I use Sonnet primarily), but it's not far off

    Qwen 3.8 27B is excellent, but it defaults to overthinking things. c16, Hacker News commenter, self-described, Aug 17, 2026. Qwen3.8 27B MLX in the author's own model harness, against Claude Sonnet.

  • Favorable
    around the same level as GPT 5.4 xhigh from march, OpenAIs leading model at that time

    Norway should buy OpenAI. NorwegianDude, Hacker News commenter, self-described, Aug 18, 2026. Qwen3.8 27B on the author's personal PC at 85 to 95 tokens per second.

  • Favorable
    the dense (with more active parameters) makes a real difference in output quality

    Ornith-1.5: From Self-Scaffolding to Self-Improvement. verdverm, Hacker News commenter, self-described, Aug 19, 2026. Qwen3.8 27B dense on an OEM DGX Spark, against mixture-of-experts models.

  • Favorable
    it's the first time I feel local coding models are actually functionally useable as daily drivers

    Unsloth Dynamic 3.0 GGUFs. latentsea, Hacker News commenter, self-described, Aug 20, 2026. Qwen3.8 27B as a local coding model through Unsloth quants.

  • Unfavorable
    it struggles to do anything useful when I'm pretty sure Claude would have probably finished the same thing in a couple of hours

    If this is true, the hyperscalers are toast. aslkalska, Hacker News commenter, self-described, Aug 21, 2026. the latest Qwen3.8 27B on the author's consumer-grade GPU.

  • Unfavorable
    I wouldn't trust those smaller models with the real work

    Why your local LLM feels dumber than it is. asats, Hacker News commenter, self-described, Aug 23, 2026. Qwen3.8 27B on an M4 Max with a 40-core GPU and 128GB at 59.5 tokens per second.

  • Favorable
    Even a 4-bit quant of Qwen3.8 27b is indistinguishable from Gemini 3.7 flash in our internal tests

    Why your local LLM feels dumber than it is. a11r, Hacker News commenter, self-described, Aug 23, 2026. a 4-bit quant of Qwen3.8 27B in the author's internal tests, on an RTX 5090 with ninfer.

  • Favorable
    It doesn't feel that far behind something like GPT 5.6 Luna

    Apple caught off guard by AI demand for Mac Mini and Mac Studio. randomblock1, Hacker News commenter, self-described, Aug 31, 2026. Qwen3.8 27B at Q4_K_M with 131K context on a 7900 XTX 24GB at 55 tokens per second with MTP.

  • Favorable
    it sits somewhere between Sonnet 5 and Opus 5 capability

    Apple caught off guard by AI demand for Mac Mini and Mac Studio. SamInTheShell, Hacker News commenter, self-described, Aug 31, 2026. Qwen3.8 27B at Q4_K_S on a Radeon 7900 XTX at 50 tokens per second, with 128k context.

  • Unfavorable
    Qwen3.8-27b didn't know that the test is about riding a bicycle, while all other larger models I tested (DeepSeek V4 Pro, Qwen3.8 Max, GLM 5.3, Kimi K3) drew the pelican riding a bicycle

    Muse Spark 1.3. umajho, Hacker News commenter, self-described, Sep 3, 2026. Qwen3.8 27B on Simon Willison's pelican bicycle test against larger models.

  • Favorable
    the quality level is spectacular for the size

    Qwen 3.8 27B available on Cerebras at 1500 tokens/s. gabri200, Hacker News commenter, self-described, Sep 4, 2026. Qwen3.8 27B under high concurrency with multi-token prediction.

  • Favorable
    it has crossed the threshold and is now a useful tool for coding and debugging

    Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses. kennywinker, Hacker News commenter, self-described, Sep 8, 2026. Qwen3.8 27B for coding and debugging, with its SWE-bench Pro score against Opus 4.6.

  • Unfavorable
    another personal benchmark problem (of a very different nature) that Qwen3.8-27B usually can't solve at all, while Opus and GLM-5.3-Flash solve it consistently

    Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses. anyfoo, Hacker News commenter, self-described, Sep 8, 2026. Qwen3.8 27B at Q5 and below on a personal benchmark problem, against Opus and GLM-5.3-Flash.

  • Unfavorable
    I have found qwen to be absolutely useless at anything but very specific, surgical code changes

    Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses. TomBombadildoze, Hacker News commenter, self-described, Sep 8, 2026. Qwen3.8 27B at 4-bit alongside a Claude Max subscription.

  • Favorable
    competitive at about Opus 4.5 to 4.7 level. That matches my experience in real tasks over the last few months

    Nine coding harnesses vs. your laptop. suprjami, Hacker News commenter, self-described, Sep 11, 2026. Qwen3.8 27B in real tasks over several months of local inference.

  • Favorable
    it's pretty good at spotting dumb stuff you miss when you proofcheck your own code

    Nine coding harnesses vs. your laptop. noir_lord, Hacker News commenter, self-described, Sep 11, 2026. Qwen3.8 27B on a 24GB VRAM GPU as a glorified linter.

  • Unfavorable
    I still haven't found a use case for Qwen3.8 27B that Qwen 3.6 35b A3b (MoE) is better at

    Best LLM for every budget, updated daily. seanmcdirmid, Hacker News commenter, self-described, Sep 24, 2026. Qwen3.8 27B against Qwen 3.6 35b A3b on the author's machine at up to 40 tokens per second.

  • Favorable
    It is currently debugging some issues in a native Mac Swift application and getting through the list of issues just fine

    Best LLM for every budget, updated daily. greggh, Hacker News commenter, self-described, Sep 24, 2026. a quant and tune of Qwen3.8 27B on an M1 Max 32GB MacBook with a Pi setup, debugging a Swift app overnight.

  • Favorable
    Qwen 3.8 has been the first local model I've tried that has consistently given usable and impressive results

    Ask HN: Which model do you use for work?. praash, Hacker News commenter, self-described, Sep 26, 2026. Qwen3.8 27B MLX running locally on an Apple M4 48GB for technical suggestions and code snippets.

  • Favorable
    Qwen3.8-27B has been there for the past two months

    Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s. stymaar, Hacker News commenter, self-described, Oct 4, 2026. Qwen3.8 27B for agentic coding against Opus 4.6, run at night on a slow GPU.

  • Favorable
    Qwen3.8-27B is surprisingly effective, somewhere in the vicinity of Opus 4.6

    Llama.cpp on a jailbroken PS5: run any GGUF model natively on the console. thombles, Hacker News commenter, self-described, Oct 17, 2026. Qwen3.8 27B on a jailbroken PS5 at low tens of tokens per second, around Opus 4.6.

  • Favorable
    Are they about on par with opus 4.6? Yes

    DeepSeek-v4.1-Flash-1017. Pragmata, Hacker News commenter, self-described, Oct 17, 2026. Qwen3.8 27B at 2bpw and 3bpw quants on long-running tasks, against Opus.

  • Unfavorable
    Qwen 3.8 27b on my 5090, barely acceptable, thinks too long, sometimes freaks out

    Kimi K3.1 scores 55 on Artificial Analysis, placed on the Pareto line. potsandpans, Hacker News commenter, self-described, Oct 24, 2026. Qwen3.8 27B on a 5090 for a PlayStation decompilation project against Opus and DeepSeek flash.

  • Favorable
    Qwen 3.8-27B is the current golden child if you have >24GB VRAM

    Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware. PaulKeeble, Hacker News commenter, self-described, Oct 17, 2026. Qwen3.8-27B as a dense local model needing more than 24GB VRAM, against mixture-of-experts models.

  • Favorable
    I just can't believe my eyes i.r.t. to how good this model is

    GLM-5.3 is now open-weight. rubin55, Hacker News commenter, self-described, Aug 29, 2026. Qwen3.8 27B tested for a week by an intensive GLM-5.2 and Opus 5 user.

  • Favorable
    It fails the simple clock test for me which Qwen3.8 27B got (nearly) right

    DeepSeek-v4-flash-vision-exp. leumon, Hacker News commenter, self-described, Aug 21, 2026. Qwen3.8 27B on a clock-reading image test against DeepSeek v4 flash vision.

  • Unfavorable
    both those models will generate an enormous amount of noise

    IBM CEO says open weight models are now good and cheap on-prem. dekhn, Hacker News commenter, self-described, Oct 26, 2026. Qwen3.8 27B on ninfer on a 5090 for a web-app coding task, against Claude and ChatGPT.

  • Unfavorable
    produces systematic Tamil agglutinative-suffix corruption on medical-domain generation that Qwen3.6-27B does not

    Tamil (agglutinative-language) quality regression vs Qwen3.6-27B, not fixed by better DeltaNet quantization protection. manoj-relusys, GitHub issue author, self-described, Sep 8, 2026. quantized Qwen3.8-27B checkpoints on vLLM for Tamil medical-domain generation.

  • Favorable
    Qwen 3.8 27B Q4_K_M is good enough for the tasks I have currently (C++20 coding on 1.3M LoC code base)

    Post by @AgiMan_WCL. Agi Man, X user, self-described, Oct 27, 2026. Qwen3.8 27B Q4_K_M running locally on 24GB VRAM at 160K context for C++20 work on a 1.3M-line codebase.

  • Favorable
    I just tried on the free qwen3.8:27b I'm running at home and it answered correctly straight away

    Post by @kjsnioctib. kjsnioctib, X user, self-described, Oct 27, 2026. the free qwen3.8:27b the author runs at home, on a single question.

Tool use and structured output 46%

Long context 75%

  • Favorable
    The 262k context means a week of feeds fits in a single prompt.

    Run Qwen3.8 27B locally: real numbers from my Mac Studio. TerminalBytes, independent tech blog, self-described, Aug 27, 2026. Qwen3.8 27B as the Ollama Q4_K_M build on a Mac Studio M3 Ultra, five timed runs against qwen3.6:27b.

  • Favorable
    It successfully completes long horizon tasks with OpenCode.

    Apple caught off guard by AI demand for Mac Mini and Mac Studio. jckahn, Hacker News commenter, self-described, Aug 31, 2026. Qwen3.8 27B as Unsloth's Q3 XXS with 128k context on a Mac mini M4 24GB with OpenCode.

  • Unfavorable
    It takes about 2 hours to fill the context

    Apple caught off guard by AI demand for Mac Mini and Mac Studio. SamInTheShell, Hacker News commenter, self-described, Aug 31, 2026. Qwen3.8 27B at Q4_K_S on a Radeon 7900 XTX at 50 tokens per second, with 128k context.

  • Favorable
    You get 260k context, ~150 token/s continuous performance, and full agentic capabilities

    Local AI Is Dead. You Are at the Funeral. diablozzq, Hacker News commenter, self-described, Aug 31, 2026. Qwen3.8 27B Quasar NVFP4 on a single 5090 with the DeepSeek Harness, multi-hour autonomous runs.

  • Favorable
    my coding agents regularly reach 200k context used without noticeable degradation

    Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses. skolos, Hacker News commenter, self-described, Sep 8, 2026. Qwen3.8 27B NVFP4 with both weights and cache quantized on an RTX 5090, coding agents to 200k context.

  • Favorable
    I've run sessions in hax where it hit that cap multiple times and compressed the context down to 15% and continued with no problem

    Pi 1.0. tingletech, Hacker News commenter, self-described, Oct 2, 2026. Qwen3.8 27B on llama-server at its default 262144 context in the hax harness.

  • Unfavorable
    the limit on context size is a huge issue for more serious work

    I'm not paying $20 for ChatGPT or Claude because a free local LLM does. pllbnk, Hacker News commenter, self-described, Oct 7, 2026. Qwen3.8 27B on an M5 Pro with 48GB at about 10 tokens per second, and on an RTX 5090.

  • Favorable
    Running locally on 24GB VRAM, 160K context, 1200tok/sec (prefill) and 60tok/sec (gen)

    Post by @AgiMan_WCL. Agi Man, X user, self-described, Oct 27, 2026. Qwen3.8 27B Q4_K_M running locally on 24GB VRAM at 160K context for C++20 work on a 1.3M-line codebase.

Developer experience 17%

  • Unfavorable
    This is a hilarious default. It's absolutely not a good way to run the model, especially on consumer hardware.

    Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things. Simon Willison, independent developer and blogger, self-described, Aug 16, 2026. Qwen3.8 27B as the 17GB Q4_K_M build in LM Studio on an M5 Max MacBook Pro and a DGX Spark.

  • Unfavorable
    Older builds fail with unknown model architecture: 'qwen35' . I hit this myself.

    Run Qwen3.8 27B locally: real numbers from my Mac Studio. TerminalBytes, independent tech blog, self-described, Aug 27, 2026. Qwen3.8 27B as the Ollama Q4_K_M build on a Mac Studio M3 Ultra, five timed runs against qwen3.6:27b.

  • Unfavorable
    the effort setting matters a lot (note that the default is xhigh ) - and it is a tricky choice, as it can overthink

    Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses. Piotr Migdał, author on the Quesma blog, self-described, Aug 26, 2026. Qwen3.8 27B in BF16 and Unsloth Q8, Q4, Q2 and 1-bit quantizations on llama.cpp, GPQA Diamond, IFBench and Terminal-Bench 2.1.

  • Favorable
    The model itself (Qwen 3.8 27B specifically) can handle being interrupted just fine and i often interrupt it to clarify things

    Five Qwen3.8-27B models match Fable 5 Using GVS5H Harness. badsectoracula, Hacker News commenter, self-described, Oct 25, 2026. Qwen3.8 27B with Unsloth UD-Q4_K_XL via llama.cpp at 131k context, interrupted mid-task.

  • Unfavorable
    only accepts `reasoning_effort ∈ {xhigh, medium, low}` and calls `raise_exception(...)` for anything else

    Qwen3.8-27B chat template rejects Claude Code's default reasoning_effort="high" via vLLM → HTTP 500. EKKO-H, GitHub issue author, self-described, Aug 19, 2026. Qwen3.8-27B-AWQ-INT4 served with vLLM behind Claude Code 2.1.235.

  • Unfavorable
    it has a real overthinking issue

    Post by @jun_song. Jun Song, X user, self-described, Aug 19, 2026. Qwen3.8-27B that the author is trying to fix with supervised fine-tuning and harness tweaks.

Cost Needs 5 accounts from 3 authors

  • Favorable
    Qwen3.8 answered the same prompts in roughly 1,000 tokens where 3.6 rambled through 2,000-3,300

    Run Qwen3.8 27B locally: real numbers from my Mac Studio. TerminalBytes, independent tech blog, self-described, Aug 27, 2026. Qwen3.8 27B as the Ollama Q4_K_M build on a Mac Studio M3 Ultra, five timed runs against qwen3.6:27b.

  • Favorable
    On OpenRouter, that would cost more than 10x what it actually costs me

    Apple caught off guard by AI demand for Mac Mini and Mac Studio. randomblock1, Hacker News commenter, self-described, Aug 31, 2026. Qwen3.8 27B at Q4_K_M with 131K context on a 7900 XTX 24GB at 55 tokens per second with MTP.

  • Unfavorable
    This is a very efficient way to burn your money, but I would not recommend it for programming

    Qwen 3.8 27B available on Cerebras at 1500 tokens/s. gpugreg, Hacker News commenter, self-described, Sep 3, 2026. Qwen3.8 27B hosted on Cerebras at 1500 tokens per second for a programming task.

  • Unfavorable
    it does a lot more testing attempts and burns a ton of tokens + compactions, so every task it takes 1.2-2x as long as 35b-a3b does

    LFM2.5-24B-A2B: Built for fast on-device agents. eyegor, Hacker News commenter, self-described, Oct 15, 2026. dense Qwen3.8 27B against Qwen3.6 35b-a3b for agentic coding under 24GB VRAM.

Speed 33%

  • Unfavorable
    Was that worth waiting 21 minutes for? Absolutely not.

    Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things. Simon Willison, independent developer and blogger, self-described, Aug 16, 2026. Qwen3.8 27B as the 17GB Q4_K_M build in LM Studio on an M5 Max MacBook Pro and a DGX Spark.

  • Unfavorable
    the new model generates at half the speed of its predecessor

    Run Qwen3.8 27B locally: real numbers from my Mac Studio. TerminalBytes, independent tech blog, self-described, Aug 27, 2026. Qwen3.8 27B as the Ollama Q4_K_M build on a Mac Studio M3 Ultra, five timed runs against qwen3.6:27b.

  • Favorable
    I'm getting about 50-60 tokens/second

    Qwen 3.8 27B. lelandbatey, Hacker News commenter, self-described, Aug 14, 2026. Qwen3.8 27B as the Unsloth 2-bit quant on a 5070ti with 16GB of VRAM at 128k context through llama.cpp.

  • Unfavorable
    this new dense model, Qwen 3.8 27B, gets ~4 tokens/second on generation

    Qwen 3.8 27B. peri-cl, Hacker News commenter, self-described, Aug 14, 2026. Qwen3.8 27B UD-Q4_K_XL on CPU against Qwen3.6-35B-A3B on a laptop.

  • Favorable
    I can run it at 85-95 tokens per second on my personal PC

    Norway should buy OpenAI. NorwegianDude, Hacker News commenter, self-described, Aug 18, 2026. Qwen3.8 27B on the author's personal PC at 85 to 95 tokens per second.

  • Favorable
    sglang, 150+ tok/s on a 5090 in ubuntu 26.04 via wsl

    Why your local LLM feels dumber than it is. IronWolve, Hacker News commenter, self-described, Aug 22, 2026. Qwen3.8 27B NVFP4 on an RTX 5090 with SGLang in opencode, medium reasoning, 96k context.

  • Unfavorable
    the new Qwen3.8-27B model is actually slower with MTP on my M4 Max

    Ornith-1.5: From Self-Scaffolding to Self-Improvement. jwr, Hacker News commenter, self-described, Aug 20, 2026. Qwen3.8 27B with multi-token prediction on an M4 Max.

  • Favorable
    Qwen3.8-27B runs at 59.5 tok/s on my M4 Max, 40-core GPU, 128 GB

    Why your local LLM feels dumber than it is. asats, Hacker News commenter, self-described, Aug 23, 2026. Qwen3.8 27B on an M4 Max with a 40-core GPU and 128GB at 59.5 tokens per second.

  • Favorable
    you can get ~800 TPS token generation (c=8) and ~140 Tokens per second single stream

    Why your local LLM feels dumber than it is. a11r, Hacker News commenter, self-described, Aug 23, 2026. a 4-bit quant of Qwen3.8 27B in the author's internal tests, on an RTX 5090 with ninfer.

  • Unfavorable
    it's difficult to run a dense model concurrently with high throughput

    Qwen 3.8 27B available on Cerebras at 1500 tokens/s. gabri200, Hacker News commenter, self-described, Sep 4, 2026. Qwen3.8 27B under high concurrency with multi-token prediction.

  • Favorable
    200 tok/s generation for single stream, or 801 tok/s generation for 8 concurrent streams

    Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses. skolos, Hacker News commenter, self-described, Sep 8, 2026. Qwen3.8 27B NVFP4 with both weights and cache quantized on an RTX 5090, coding agents to 200k context.

  • Unfavorable
    the t/s is so low for me on my old M1 Max 64GB that I hope others are getting use out of it

    Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses. bellowsgulch, Hacker News commenter, self-described, Sep 8, 2026. Qwen3.8 27B on an old M1 Max 64GB.

  • Favorable
    I can run Qwen 3.8 27B (which is a REALLY capable model) in the background coding for me while I'm simultaneously browsing the web and playing VALORANT without any performance impact

    Mistral X Mozilla: Private, Multilingual AI Browsing. redox99, Hacker News commenter, self-described, Sep 16, 2026. Qwen3.8 27B coding in the background on a six-year-old GPU while gaming and browsing.

  • Unfavorable
    27B thinks so much that it was honestly too painful

    Xiaomi MiMo v2.6. girvo, Hacker News commenter, self-described, Sep 21, 2026. Qwen3.8 27B against Qwen 3.8 Flash Next on a GB10 box with the Pi harness.

  • Unfavorable
    I can get at most 40 tokens/second with 27B, but I get around 90 tokens/second with the MoE

    Best LLM for every budget, updated daily. seanmcdirmid, Hacker News commenter, self-described, Sep 24, 2026. Qwen3.8 27B against Qwen 3.6 35b A3b on the author's machine at up to 40 tokens per second.

  • Unfavorable
    they take much more time than Opus due to the compounding effects of my slow GPU and Qwen's verbosity

    Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s. stymaar, Hacker News commenter, self-described, Oct 4, 2026. Qwen3.8 27B for agentic coding against Opus 4.6, run at night on a slow GPU.

  • Unfavorable
    I am getting ~10 tok/s in M5 Pro with 48 GB with Qwen 3.8 27B. It's workable but very annoying

    I'm not paying $20 for ChatGPT or Claude because a free local LLM does. pllbnk, Hacker News commenter, self-described, Oct 7, 2026. Qwen3.8 27B on an M5 Pro with 48GB at about 10 tokens per second, and on an RTX 5090.

  • Unfavorable
    anything moderately involved requires xhigh reasoning and it will spend literally tens of thousands of tokens turning over a problem

    Llama.cpp on a jailbroken PS5: run any GGUF model natively on the console. thombles, Hacker News commenter, self-described, Oct 17, 2026. Qwen3.8 27B on a jailbroken PS5 at low tens of tokens per second, around Opus 4.6.

  • Unfavorable
    because its a dense model it is pretty slow to run and requires a lot of hardware

    Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware. PaulKeeble, Hacker News commenter, self-described, Oct 17, 2026. Qwen3.8-27B as a dense local model needing more than 24GB VRAM, against mixture-of-experts models.

  • Unfavorable
    the MTP variants of Qwen3.8-27B as consistently slower than the same quantization without MTP heads, on Apple Silicon

    Qwen3.8-27B MTP variants are 2x slower than non-MTP on Apple Silicon - is this expected?. otsoa-dubief, GitHub issue author, self-described, Aug 15, 2026. Qwen3.8-27B multi-token-prediction quants against the same quant without heads on Apple Silicon.

  • Unfavorable
    qwen3.8 27b on my hardware runs slow as fuck so i wouldnt use any of these

    Post by @G4mi_. Gami, X user, self-described, Oct 27, 2026. Qwen3.8 27B on the author's own hardware.

Reliability Only concerns on record