Skip to main content
Fireworks AIProduct page

Fireworks AI Signals

What teams building AI features on Fireworks AI report, with every finding tied to its evidence.

Read the findings

What the evidence says

Whose sources count

Practitioner sentiment, all practitioners

−17 Mixed from 47 sources that take a side

How this figure is made

Mixed overall signal, net 17 points negative

From 42 authors in 43 discussions. Run a pilot on your own prompts at expected concurrency to measure time to first token, output speed, error rate and cost per million tokens.

Analyzed
76 total sources
Top praise
Low Per-Token Bills (8 of 10 sources) See praise evidence
Top concern
Price Rises and Cached Tokens (9 of 12 sources) See concern evidence

Jan 2025 to Oct 2026.

47 sources from practitioners, by month and sentiment.

Across 47 counted sources the read leans critical: 14 positive, 13 mixed and 20 negative. Speed draws praise in 8 sources and low per-token bills in 8, while price rises, cached-token costs and zero-retention premiums are raised in 9 and API compatibility gaps in 6.

Positive, mixed and negative sources

The score counts them as 11 positive, 13 mixed and 18 negative votes, one per author. See How the numbers work.

See the evidence

Below: 3 praise themes, 3 concerns and 1 disagreement.

Findings

Counted for All practitioners, from 47 sources. Change group

What practitioners praise

Every capability practitioners describe positively, ranked by how many sources say so. Open one to read who said it.

  1. Low Per-Token Bills8 of 10 sources on it praise itPaying per token for open models on Fireworks costs less than the frontier subscription or model the author compares it with.

    Counted for All practitioners.

    Positive, 8

    1. Positive

      kelnos, Core developer at Xfce

      Hacker News comment by kelnosHacker News comment

      Coding with GLM 5.3 Flash from Fireworks costs them US$1 to US$2 per day, so at most US$60 a month, against the US$100 a month Claude plan they used before.

      Published Oct 9, 2026self-described: Individual coding use, after a US$100 a month Claude plan

    2. Mixed

      ar27111994, Reddit poster

      Self-hosting Qwen3.8-Flash-Next (or a smaller alternative) for a heavy multi-agent Hermes setup on $10k of AWS credits, looking for options I've missedReddit post

      Ran roughly 9 to 15B tokens a month on Fireworks with a 90 to 98% cache hit rate, which they call a big part of the economics. They are moving to self-hosting only because their one budget is US$10k of AWS Activate credits.

      Published Oct 4, 2026self-described: A multi-agent orchestrator moving to self-hosting paid with AWS Activate credits

    3. Positive

      mark_l_watson, author and AI practitioner

      Hacker News comment by mark_l_watson: Experimenting with open models instead of buying a bigger local machineHacker News comment

      Pays as they go on Fireworks for open-model inference, and however much they use it their monthly bill is between US$10 and US$40, much faster than a home rig.

      Published Aug 25, 2026self-described: Experimenting with open models and choosing not to buy a bigger local machine

    4. Positive

      vijucat, Algorithmic trading developer

      Hacker News comment by vijucatHacker News comment

      On Fireworks a cache hit costs only 20% of the input price, and uncached DSV4-0731 is US$0.14 per million tokens, so whole re-architecture projects cost them a few dollars. They found GitHub Copilot poor at using the cache.

      Published Aug 8, 2026self-described: Long agentic coding projects with prompt caching

    5. Positive

      drewnick, Hacker News commenter

      Hacker News comment by drewnickHacker News comment

      Moved one of their company's daily workflows to Kimi K3 on Fireworks, where it replaces two sales assistant positions for about US$50 a day for 14 million tokens in and out.

      Published Jul 29, 2026self-described: A daily sales workflow moved to an open model

    6. Positive

      mark_l_watson, author and AI practitioner

      Hacker News comment by mark_l_watson: Daily work moved from Opus to an open modelHacker News comment

      After a month of paying heavily for Opus, they get a lot done on GLM 5.2 running on Fireworks for about 15% of the money, and say everyone should run their own evals on their own work.

      Published Jul 7, 2026self-described: Daily work moved from Opus to an open model

    7. Positive

      Hussain Gagan, PeerSpot reviewer

      Gaining faster, flexible AI workflows has made our team ship reliable features with confidencePeerSpot review

      After six to eight months, his team sees about 20 to 30% lower latency on some endpoints and 15 to 25% savings depending on the model, which he calls meaningful. He finds documentation for advanced configs unclear and model behavior hard to debug. PeerSpot gave him a gift card for the review.

      Published Apr 20, 2026self-described: Summarization and internal-search APIs moved to open models on Fireworks

    8. Positive

      vuldin, Hacker News commenter

      Hacker News comment by vuldinHacker News comment

      Switched from Claude Code with Opus 4.5 to OpenCode with Kimi K2.5 from Fireworks. They never hit time-based limits now, and their spend went from well over US$100 per month to less than US$50 per month.

      Published Feb 18, 2026self-described: Moving from a coding subscription to an open model

    Mixed or negative, 2

    1. Mixed

      rapind, Hacker News commenter

      Hacker News comment by rapindHacker News comment

      Ran DeepSeek V4.1 Flash on Fireworks with zero data retention and racked up more than US$100 in about 2 to 3 days, which they estimated at US$400 or more a month. Fireworks has since raised its prices and got slower, they say.

      Published Sep 29, 2026self-described: Agentic coding on a zero-data-retention host, with a 98 to 99% cache rate

    2. Negative

      WinstonSmith84, Hacker News commenter

      Hacker News comment by WinstonSmith84Hacker News comment

      Replacing a subscription model with DeepSeek V4.1 Flash burned about US$7 of tokens on Fireworks in a single day, more than their OpenAI subscription costs.

      Published Sep 23, 2026self-described: Replacing a subscription coding model with an open model

    GILD's fact checks, which don't count as sources

    Checked Nov 1, 2026.Fireworks Serverless text and vision prices are input / cached input / output in USD per 1M tokens. Ember-1 Standard is US$3.00 / US$0.30 / US$15.00. Serverless pricing (text)

    Checked Nov 1, 2026.Fireworks embeddings are billed per 1M input tokens, at US$0.008 for base models up to 150M parameters. Serverless pricing (text)

    Checked Nov 1, 2026.Fireworks AI uses a usage-based pre-paid billing system: per token for serverless inference, per GPU usage time for on-demand deployments, and per training token for Managed Training. Cost FAQ

    Checked Nov 1, 2026.The Fireworks Batch API processes large volumes of requests asynchronously at 50% off Serverless per-token prices. Not all models support the Batch API. Batch inference

    Open them in the record
  2. Serving Speed8 of 12 sources on it praise itOpen models on Fireworks respond fast enough for interactive work.

    Counted for All practitioners.

    Positive, 8

    1. Positive

      eli, Co-Founder of Industry Dive

      Hacker News comment by eliHacker News comment

      Calls Fireworks solid, and says its fast variant suits buyers who care more about speed than cost, at about 2x the cost.

      Published Jun 30, 2026self-described: Picking a serving tier

    2. Mixed

      Ekin Koc (eknkc), kovan studio

      Hacker News comment by eknkcHacker News comment

      Muse Spark Open on Fireworks felt weaker than GLM 5.3 but was really fast, on the single large prompt they tried.

      Published Oct 15, 2026self-described: Trying a new open-weight release on launch day

    3. Positive

      mark_l_watson, author and AI practitioner

      Hacker News comment by mark_l_watson: Experimenting with open models instead of buying a bigger local machineHacker News comment

      Pays as they go on Fireworks for open-model inference, and however much they use it their monthly bill is between US$10 and US$40, much faster than a home rig.

      Published Aug 25, 2026self-described: Experimenting with open models and choosing not to buy a bigger local machine

    4. Mixed

      pimeys, twin.so

      Fireworks providerGitHub issue

      Asks the Maki coding agent to support Fireworks natively so his whole team can use the fast Kimi K3, and notes that Fireworks carries fewer models than OpenRouter.

      Published Jul 30, 2026self-described: A team that wants Fireworks as a native provider in its coding agent

    5. Positive

      Hussain Gagan, PeerSpot reviewer

      Gaining faster, flexible AI workflows has made our team ship reliable features with confidencePeerSpot review

      After six to eight months, his team sees about 20 to 30% lower latency on some endpoints and 15 to 25% savings depending on the model, which he calls meaningful. He finds documentation for advanced configs unclear and model behavior hard to debug. PeerSpot gave him a gift card for the review.

      Published Apr 20, 2026self-described: Summarization and internal-search APIs moved to open models on Fireworks

    6. Positive

      eteitaxiv, Reddit poster

      Reddit post by eteitaxivReddit post

      Uses Fireworks almost exclusively because it is very fast and, in the poster's words, one of the only hosts that offers zero data retention with proper certification. Buyers can check Fireworks' own security pages for the certifications.

      Published Feb 15, 2026self-described: Testing chat presets on open models through Fireworks and official APIs

    7. Mixed

      extr, Hacker News commenter

      Hacker News comment by extrHacker News comment

      The speed through Fireworks is great, they say, but the open model they tried had to go back and fix type errors and broken tests far more often than Opus 4.5.

      Published Jan 31, 2026self-described: Implementing specs on an enterprise TypeScript service

    8. Positive

      keithwhor, building Superuser

      Hacker News comment by keithwhorHacker News comment

      Some OpenRouter providers gave them mangled outputs. Fireworks serverless was their best experience: fast, performant and less expensive. They had not tried custom deployments.

      Published Mar 7, 2025self-described: A startup that wants a US-hosted open model

    Mixed or negative, 4

    1. Negative

      switchbak, Hacker News commenter

      Hacker News comment by switchbak: Trying Fireworks for speed against DeepSeek's own APIHacker News comment

      Tried Fireworks for its advertised speed and found it mostly worse than DeepSeek's own service.

      Published Oct 9, 2026self-described: Trying Fireworks for speed against DeepSeek's own API

    2. Negative

      switchbak, Hacker News commenter

      Hacker News comment by switchbak: Comparing Fireworks with Baseten for open-model speedHacker News comment

      Finds Fireworks less strong lately and is having a better time on Baseten, where they see the speeds Fireworks advertises, though they have only just started there.

      Published Oct 26, 2026self-described: Comparing Fireworks with Baseten for open-model speed

    3. Mixed

      rapind, Hacker News commenter

      Hacker News comment by rapindHacker News comment

      Ran DeepSeek V4.1 Flash on Fireworks with zero data retention and racked up more than US$100 in about 2 to 3 days, which they estimated at US$400 or more a month. Fireworks has since raised its prices and got slower, they say.

      Published Sep 29, 2026self-described: Agentic coding on a zero-data-retention host, with a 98 to 99% cache rate

    4. Mixed

      MarcusAurelius68, Reddit poster

      When a 3060 is faster than a 5060tiReddit post

      Embedding one publication took 9 minutes on their local RTX 3060. Fireworks.ai and OpenRouter were about the same or slightly slower, with chunks sent one at a time, which they call possibly an area for improvement.

      Published Jul 16, 2026self-described: Embedding publications for a RAG pipeline, one chunk at a time

    GILD's fact checks, which don't count as sources

    Checked Nov 1, 2026.Fireworks deploys models on dedicated GPUs with fast autoscaling and minimal cold starts. Introduction

    Checked Nov 1, 2026.On-demand deployments give you dedicated GPUs for your models. On-demand deployments

    Checked Nov 1, 2026.Fireworks Serverless Inference is multi-tenant inference for popular open models running on Fireworks-managed infrastructure. Serverless overview

    Open them in the record
  3. Moving Work to Open Models5 of 7 sources on it praise itTeams move coding, review and business workflows from closed models to open models served by Fireworks, and keep their existing clients and frameworks.

    Counted for All practitioners.

    Positive, 5

    1. Positive

      verdverm, Hacker News commenter

      Hacker News comment by verdverm: A development team moving from Claude to open modelsHacker News comment

      The commenter's team fully rolled out its custom PR review on glm-5.3-flash, and is moving developers who are still on Claude toward Fireworks and OpenCode.

      Published Oct 2, 2026self-described: A development team moving from Claude to open models

    2. Positive

      drewnick, Hacker News commenter

      Hacker News comment by drewnickHacker News comment

      Moved one of their company's daily workflows to Kimi K3 on Fireworks, where it replaces two sales assistant positions for about US$50 a day for 14 million tokens in and out.

      Published Jul 29, 2026self-described: A daily sales workflow moved to an open model

    3. Positive

      mark_l_watson, author and AI practitioner

      Hacker News comment by mark_l_watson: Daily work moved from Opus to an open modelHacker News comment

      After a month of paying heavily for Opus, they get a lot done on GLM 5.2 running on Fireworks for about 15% of the money, and say everyone should run their own evals on their own work.

      Published Jul 7, 2026self-described: Daily work moved from Opus to an open model

    4. Positive

      Hussain Gagan, PeerSpot reviewer

      Gaining faster, flexible AI workflows has made our team ship reliable features with confidencePeerSpot review

      After six to eight months, his team sees about 20 to 30% lower latency on some endpoints and 15 to 25% savings depending on the model, which he calls meaningful. He finds documentation for advanced configs unclear and model behavior hard to debug. PeerSpot gave him a gift card for the review.

      Published Apr 20, 2026self-described: Summarization and internal-search APIs moved to open models on Fireworks

    5. Positive

      vuldin, Hacker News commenter

      Hacker News comment by vuldinHacker News comment

      Switched from Claude Code with Opus 4.5 to OpenCode with Kimi K2.5 from Fireworks. They never hit time-based limits now, and their spend went from well over US$100 per month to less than US$50 per month.

      Published Feb 18, 2026self-described: Moving from a coding subscription to an open model

    Mixed or negative, 2

    1. Mixed

      __jl__, Hacker News commenter

      Hacker News comment by __jl__Hacker News comment

      The commenter's team used open-weight models through Fireworks for a long time, Kimi K2.5 among them, and calls Fireworks a great provider that still had issues here and there. For its production app, OpenAI is now the best provider in the commenter's view.

      Published Jul 30, 2026self-described: A production app that ran open-weight models on Fireworks

    2. Mixed

      extr, Hacker News commenter

      Hacker News comment by extrHacker News comment

      The speed through Fireworks is great, they say, but the open model they tried had to go back and fix type errors and broken tests far more often than Opus 4.5.

      Published Jan 31, 2026self-described: Implementing specs on an enterprise TypeScript service

    GILD's fact checks, which don't count as sources

    Checked Nov 1, 2026.You can use the OpenAI Python client library to interact with Fireworks. OpenAI compatibility

    Checked Nov 1, 2026.Fireworks documents building production AI agents on Fireworks with LangChain, LlamaIndex, CrewAI, AutoGen, and other open-source frameworks using the OpenAI-compatible API. Agent frameworks

    Checked Nov 1, 2026.Fireworks lists LiteLLM (native fireworks_ai provider) and Helicone (Fireworks proxy and provider integration) among the gateways that connect to your Fireworks account. LLM gateways

    Checked Nov 1, 2026.Fireworks AI is a first-party inference provider inside Microsoft Foundry, with usage billed through Azure and counting toward your Microsoft Azure Consumption Commitment (MACC). Microsoft Foundry

    Checked Nov 1, 2026.On Microsoft Foundry, Fireworks PayGo deployments run in the US Data Zone only. Foundry PTU (Provisioned Throughput) deployments provide dedicated GPU capacity reserved for your workload, with global region availability. Microsoft Foundry

    Checked Nov 1, 2026.Fireworks' `auto` model ID picks a model for each new user turn. Closed models run on your own Anthropic or OpenAI account. To choose a model yourself, replace `auto` with its ID. When Fireworks' CLI connects Claude Code, Auto is the default model in `/model`. Nexus Auto

    Checked Nov 1, 2026.Before you begin, create an API key in the Fireworks dashboard. Quickstart

    Checked Nov 1, 2026.Fireworks is self-serve. Contact Fireworks for enterprise deployments. Pricing

    Open them in the record

What concerns them

Every concern practitioners raise, most raised first.

  1. Price Rises and Cached Tokens9 of 12 sources on it raise it as a concernList prices rose, cached-token rates add up, or zero-retention hosting costs more than the model maker's own API.

    Counted for All practitioners.

    Raised as a concern, 9

    1. Negative

      pkulak, Engineer at Vevo

      Hacker News comment by pkulakHacker News comment

      Fireworks emailed them that it was raising the price by about 30%, and they expect to switch hosts.

      Published Sep 22, 2026self-described: Paying for a popular open model through Fireworks

    2. Mixed

      rapind, Hacker News commenter

      Hacker News comment by rapindHacker News comment

      Ran DeepSeek V4.1 Flash on Fireworks with zero data retention and racked up more than US$100 in about 2 to 3 days, which they estimated at US$400 or more a month. Fireworks has since raised its prices and got slower, they say.

      Published Sep 29, 2026self-described: Agentic coding on a zero-data-retention host, with a 98 to 99% cache rate

    3. Negative

      vardalab, Hacker News commenter

      Hacker News comment by vardalabHacker News comment

      Cache reads for GLM 5.3 Flash on Fireworks' API cost much more than for DeepSeek, so a simple run with an advisor model cost them US$3 in oversight.

      Published Sep 24, 2026self-described: An agent setup with a stronger model as an advisor over an open worker model

    4. Negative

      WinstonSmith84, Hacker News commenter

      Hacker News comment by WinstonSmith84Hacker News comment

      Replacing a subscription model with DeepSeek V4.1 Flash burned about US$7 of tokens on Fireworks in a single day, more than their OpenAI subscription costs.

      Published Sep 23, 2026self-described: Replacing a subscription coding model with an open model

    5. Mixed

      ofisboy, Building frontier.legal

      Hacker News comment by ofisboyHacker News comment

      Uses Fireworks for a production app at low volume. The one complaint is that serverless reranking offers a single model, Qwen 8b, and it is quite expensive.

      Published Sep 11, 2026self-described: A production app at low volume

    6. Negative

      bwfan123, Hacker News commenter

      Hacker News comment by bwfan123: Buying zero-data-retention inference directly from FireworksHacker News comment

      Zero-data-retention hosts such as Fireworks do not disclose caching and tokens clearly, they say, so those hosts end up far more expensive than DeepSeek's own API.

      Published Jul 31, 2026self-described: Buying zero-data-retention inference directly from Fireworks

    7. Negative

      bwfan123, Hacker News commenter

      Hacker News comment by bwfan123: Zero-data-retention access to DeepSeek modelsHacker News comment

      Zero-data-retention providers of DeepSeek models, OpenRouter with ZDR guardrails and Fireworks among them, cost them 3x to 5x more than DeepSeek directly, which they call the price of ZDR.

      Published Jul 19, 2026self-described: Zero-data-retention access to DeepSeek models

    8. Mixed

      Chaheti Jha, PeerSpot reviewer

      Model testing has become faster and fine-tuning now supports flexible customizationPeerSpot review

      Values the catalog of 300-plus models and the fine-tuning options, but finds Fireworks very much more expensive than Groq, which she generally uses, with no free tier and weak image and video generation. PeerSpot gave her a gift card for the review.

      Published May 5, 2026self-described: One month of model testing and fine-tuning

    9. Negative

      cunningjames, Reddit poster

      Using Cline with a third-party hosted DeepSeek-V3, spits out junk when editing filesReddit post

      DeepSeek-V3 hosted on Fireworks worked for a while in Cline at a price they call certainly higher, then began writing part of each file followed by random characters and a truncation. They ask whether DeepSeek or the host is at fault.

      Published Jan 31, 2025self-described: Cline with DeepSeek-V3 hosted on Fireworks

    Discussed without concern, 3

    1. Mixed

      ar27111994, Reddit poster

      Self-hosting Qwen3.8-Flash-Next (or a smaller alternative) for a heavy multi-agent Hermes setup on $10k of AWS credits, looking for options I've missedReddit post

      Ran roughly 9 to 15B tokens a month on Fireworks with a 90 to 98% cache hit rate, which they call a big part of the economics. They are moving to self-hosting only because their one budget is US$10k of AWS Activate credits.

      Published Oct 4, 2026self-described: A multi-agent orchestrator moving to self-hosting paid with AWS Activate credits

    2. Positive

      vijucat, Algorithmic trading developer

      Hacker News comment by vijucatHacker News comment

      On Fireworks a cache hit costs only 20% of the input price, and uncached DSV4-0731 is US$0.14 per million tokens, so whole re-architecture projects cost them a few dollars. They found GitHub Copilot poor at using the cache.

      Published Aug 8, 2026self-described: Long agentic coding projects with prompt caching

    3. Mixed

      Héctor Luaces (hluaces), Infrastructure engineer at Igalia

      [Bug]: fireworks session id is reset to a random string when header is not provided causing caching to not workGitHub issue

      He traces the cause to LiteLLM, which generated a random session id per request so cached tokens stayed at 0. Calling Fireworks without a per-request id returned 9007 cached tokens on the second and third identical calls.

      Published Aug 3, 2026self-described: Calling Fireworks through a LiteLLM proxy and checking cache hits

    Open them in the record
  2. API Compatibility GapsAll 6 sources on it raise it as a concernRequests, tools and clients that work elsewhere fail validation on Fireworks' compatible endpoints or its CLI.

    Counted for All practitioners.

    1. Negative

      HareeshBahuleyan, Mozilla.ai, machine learning and NLP

      invalid_request_error when using models from Fireworks AIGitHub issue

      A plain completion with gpt-oss-120b on Fireworks works, but after a few agent steps Fireworks returns a 400 that rejects OpenAI message fields such as refusal and annotations. He is not sure whether any-agent or any-llm is at fault.

      Published Aug 26, 2025self-described: Running the any-agent framework's tool loop on Fireworks

    2. Negative

      Mikael Koskinen, .NET and Azure developer (adafy.com)

      Claude Code subagents with the new feature "Claude Code keeps its own tier models with the Fireworks catalog added"GitHub issue

      A Fireworks CLI release stopped replacing Claude Code's built-in model slots, so his subagents could no longer run on custom Fireworks models. A collaborator on Fireworks' repository replied that the override is fixed in release 0.9.8.

      Published Sep 22, 2026self-described: Running Claude Code subagents on Fireworks models through Fireworks' CLI

    3. Negative

      Duc Nguyen, Senior Software Developer at Wealthsimple

      GLM models on Fireworks fail with prompt_cache_retention errorGitHub issue

      GLM-5.2 on Fireworks returned a 400 for the prompt_cache_retention parameter his coding agent sends by default, while other Fireworks models such as DeepSeek and Kimi worked with caching.

      Published Aug 5, 2026self-described: A coding agent that sends OpenAI prompt-cache fields

    4. Negative

      Nyasha Chiroro, full stack web developer

      Error trying to send messages via Fireworks model providerGitHub issue

      Every message sent through Fireworks as the provider returned a 400 JSON Schema not supported error in the author's coding agent. The agent's maintainer replied that the error could not be reproduced.

      Published Jun 3, 2026self-described: A coding agent with conditional tool schemas

    5. Negative

      Hector Garcia (hectahertz), Buffer

      Fireworks.ai models fail with 400 on tool-enabled requests via anthropic-messages APIGitHub issue

      Every Fireworks model called through the anthropic-messages API returns a 400 when the request includes tools, because the endpoint does not accept eager_input_streaming or cache_control on tools.

      Published Apr 27, 2026self-described: An Anthropic-format coding agent pointed at Fireworks

    6. Negative

      scosman, Building kiln.tech

      [Bug] Fireworks fine-tune error "Dataset ID: first character of id must not be a digit"GitHub issue

      Fireworks' fine-tuning API rejects dataset IDs whose first character is a digit, so Kiln users retry until a UUID starts with a letter. He wants Fireworks to accept UUIDs, and if it will not, plans to prepend KILN-.

      Published Apr 19, 2025self-described: Kiln, a fine-tuning and evaluation app that calls Fireworks

    GILD's fact checks, which don't count as sources

    Checked Nov 1, 2026.Fireworks Anthropic compatibility serves only POST /v1/messages and does not support Anthropic server side tools. Anthropic compatibility

    Open them in the record
  3. Varies by Model and Day4 of 5 sources on it raise it as a concernSpeed and output quality depend on the model and the day, and teams report issues here and there.

    Counted for All practitioners.

    Raised as a concern, 4

    1. Mixed

      svachalek, Software engineering manager

      Hacker News comment by svachalekHacker News comment

      Fireworks can be great, they say, but it depends on the model and on the day.

      Published Sep 11, 2026self-described: Choosing hosts for open-weight models

    2. Mixed

      verdverm, Hacker News commenter

      Hacker News comment by verdverm: A team that selected Fireworks as its open-model vendorHacker News comment

      When the commenter's team selected Fireworks, new models arrived in 1 to 2 days. In the last few months it has not kept up and asks users to request models in Discord, and its Ember-1 model, Kimi K3 tuned to think less, hallucinates badly in the commenter's use.

      Published Oct 19, 2026self-described: A team that selected Fireworks as its open-model vendor

    3. Mixed

      __jl__, Hacker News commenter

      Hacker News comment by __jl__Hacker News comment

      The commenter's team used open-weight models through Fireworks for a long time, Kimi K2.5 among them, and calls Fireworks a great provider that still had issues here and there. For its production app, OpenAI is now the best provider in the commenter's view.

      Published Jul 30, 2026self-described: A production app that ran open-weight models on Fireworks

    4. Negative

      cunningjames, Reddit poster

      Using Cline with a third-party hosted DeepSeek-V3, spits out junk when editing filesReddit post

      DeepSeek-V3 hosted on Fireworks worked for a while in Cline at a price they call certainly higher, then began writing part of each file followed by random characters and a truncation. They ask whether DeepSeek or the host is at fault.

      Published Jan 31, 2025self-described: Cline with DeepSeek-V3 hosted on Fireworks

    Discussed without concern, 1

    1. Positive

      keithwhor, building Superuser

      Hacker News comment by keithwhorHacker News comment

      Some OpenRouter providers gave them mangled outputs. Fireworks serverless was their best experience: fast, performant and less expensive. They had not tried custom deployments.

      Published Mar 7, 2025self-described: A startup that wants a US-hosted open model

    Open them in the record

Where the evidence agrees

Statements this group's sources agree on, then the ones close to agreement.

For all practitioners, no statement reaches 5 sources from 3 authors.

Where it disagrees

Questions the sources split on, with the count on each side.

  • Evidence remains split

    Does Fireworks cost practitioners less than what they compared it with?

    • Costs less, 7 sources

      Developers moving off frontier subscriptions or models report lower bills on open models through Fireworks.

    • Costs more, 7 sources

      Others found daily spend above their subscription, a premium over the model maker's own API, or a fine-tuning bill far above GPU rental.

The evidence

Whose experience this is

The three groups, then every practitioner by the context they work in, and how emphasis changes with scale.

Three groups drawn from the same sources. Choose a group to select whose experience counts. Within that group, each author counts once in the practitioner sentiment, with their latest view. Signals open on All practitioners. Teams and leaders have 16 accounts; GILD prints their sentiment from 20.

  • Teams and technical leaders

    16 sources. Small teams and named companies describe production use, model launches and capacity on Fireworks.

  • Individual builders

    31 sources. Individual developers describe their own spend and speed on open models served by Fireworks.

  • All practitioners

    47 sources. Together, these accounts combine individual developers' bills with a few teams' production accounts.

47 sources from practitioners, by the context they work in. Filled rows count for this group.

  1. Individual builderDevelopers building on the platform on their own31 sources, counted here

    Emphasizes Daily and monthly spend against a subscription, and whether the advertised speed shows up

  2. Small product team2 to 10 engineers, or an unnamed startup team12 sources, counted here

    Emphasizes Team rollouts on open models, model launches and catalog gaps

  3. Product or platform team11 to 50 engineers, or a named company's team2 sources, counted here

    Emphasizes GitLab's capacity redirect and request validation errors in an agent framework

  4. Technical leadershipLeaders setting model and platform choices across teams2 sources, counted here

    Emphasizes Founders on serverless quality and a faster tier at about 2x the cost

  5. Organization-wideEnterprise or company-wide AI platformNo sources from this setting yet

What changes with scale. Individual developers mostly weigh daily spend against a subscription, and split on whether Fireworks costs less. Teams and named companies report production use, launches and catalog gaps, and GitLab's record shows a new on-demand deployment that would not scale.

Where the sources come from

Where this group's sources were published.

Counted for All practitioners, from 47 sources. Change group

76 sources for this group, 29 of them documentation

  • Developer Communities44, 58% of 76 sources
  • Product Documentation29, 38% of 76 sources
  • Review Sites3, 4% of 76 sources

Earlier evidence remains when it still applies. A bug report closed without a stated fix is marked Needs re-verification.

See sources by origin

The record

Who each source is, how it knows, whether it still applies and what it counts toward.

Newest first, on its own page. Every count above opens its sources there.

Read all 76 sources

About this read

What changed, and how we count

What changed

  • Nov 1, 2026: adds 7 GitHub issues, 4 Reddit posts and 3 PeerSpot reviews that a different model checked, including a US$40,000 AWS Marketplace charge one reviewer reports and request validation errors on Fireworks' compatible APIs.
  • Nov 1, 2026: first edition, built from 33 practitioner records and 32 documentation statements that a different model checked against their stored pages.
  • Every account links to its public source and names its author and publication date.
  • A finding shows only when at least 5 sources from 3 different authors discuss it. A disagreement also requires at least 2 sources on each side.
  • Checks the pricing, limits, region and security findings against Fireworks' documentation stored on the review date.

Method

A running review of what teams building AI features on Fireworks AI report, from individual builders to technical leaders. Each finding links to its sources, and each conflict shows both sides.

Counts cover only the sources GILD reviewed.

Who counts

Teams and technical leaders
Small product team, Product or platform team, Technical leadership, Organization-wide
Individual builders
Individual builder
All practitioners
Everyone in the teams and leaders and individual builders groups

Choose a group to select whose experience counts. Within that group, each author counts once in the practitioner sentiment, with their latest view. GILD files each source in a group from how the account describes its author. A role or employer marked self-described is what the source states about themselves. Documentation checks facts for every group and never counts toward sentiment.

How the numbers work

  • Overall signal. Positive votes minus negative ones, per 100 votes from the sources that take a side. Each author counts once, with their latest view, whoever they are and wherever they posted. The older weighted reading, where accounts GILD matched to a named person count three times when GILD lists the author among technical leaders and twice otherwise, is shown beside it as a second view. Where an author wrote more than one source, the line beside the split gives the votes, one per author, so the score follows from it. When one author's sources from the same day disagree, that author counts as mixed, and a half point rounds away from zero. Sources with no opinion are listed and not counted. Strong positive is +50 or higher, positive +20 to +49, mixed −19 to +19, negative −20 to −49, strong negative −50 or lower.
  • Weighted reading, second view. Counted the older way, where an account GILD matched to a named person counts two or three times, all practitioners read −13 mixed (method sentiment/S1). GILD printed that number as "GILD score" until October 29, 2026; most of those matches rest on a name or a self-described role.
  • Unscored. A group gets a score only with 20 sources from 15 authors. Below that, one source moves the score about 5 points, so the read says "Unscored until 20 sources from 15 authors" and shows its counts.
  • Independence. A source whose author works for the vendor, for one of its partners or programs, or for a competitor, and a page the vendor published or ran, is listed with that tie and counts toward nothing. One author counts on at most two records per product. The line under the score says how many authors and how many separate discussions the counted sources come from: comments in one thread are one discussion.
  • Praise and concerns. How many of the sources that discuss a theme praise it or raise it.
  • Consensus. A statement joins when 20% of the group's sources support it and fewer contradict it. Contradicting sources are counted separately. Below that line a statement shows as building. At 40% or more it reads Strong agreement; below, Recurring point.
  • Disagreements. The count on each side. GILD names no winner and gives no share of all users.
  • Minimum evidence. A finding needs 5 sources from 3 authors; a disagreement also needs 2 on each side. Smaller counts stay in the table below, marked under the minimum.

Labels on a source

A source's date and whether it still applies are separate: a recent post can describe an old version, and an old one can still hold.

Documented
Stated in official documentation.
Reported
A practitioner's account of their own use.
Measured
Backed by the source's own measurements, such as a comparative test or tracked team metrics.
GILD analysis
GILD's check of a claim against official documentation.
Current
Describes the product as it is now.
Still applicable
Older, and the behavior or lesson still holds.
Needs a recheck
May depend on a plan, version or setting that changed since it was written. It keeps counting until GILD rechecks it, since the rest of the account usually holds, and its card says so.
Out of date
No longer accurate. Kept in the record; counts toward nothing current.

Flagged for a recheck, of the sources that take a side: teams and leaders 1 of 16, individual builders 0 of 31, all practitioners 1 of 47.

Limits

  • These figures describe reviewed sources; they do not represent all Fireworks customers. Hacker News supplies 28 of the 47 counted sources, GitHub issues 11, Reddit posts 4, PeerSpot reviews 3 and a GitLab merge request 1, and a bug report counts as negative. Two of the three PeerSpot reviewers received a PeerSpot gift card for their review. Cursor's May 2024 post and Sourcegraph's December 2023 podcast appear as Operator Insights only, since Signals counts sources from 2025 on.
  • Sources run from January 2025 to October 2026 and most name the model they ran, among them GLM-5.3-Flash, Kimi K3, DeepSeek V4.1 Flash and GLM-5.2. Spend figures are the authors' own bills; Fireworks' list prices are in the Review.

Every count, and the sources behind it

17 findings for 3 groups

Each row opens its sources for the group you are reading.

FindingTeams and technical leadersIndividual buildersAll practitionersSources
Current readCurrent sentiment6 positive, 5 mixed, 5 negative8 positive, 8 mixed, 15 negative14 positive, 13 mixed, 20 negativeSources
PraiseLow Per-Token Bills2 of 2 positive (under the minimum)6 of 8 positive8 of 10 positiveSources
PraiseServing Speed4 of 4 positive (under the minimum)4 of 8 positive8 of 12 positiveSources
PraiseNew Models Quickly1 of 2 positive (under the minimum)0 of 0 positive (under the minimum)1 of 2 positive (under the minimum)Sources
PraiseMoving Work to Open Models3 of 4 positive (under the minimum)2 of 3 positive (under the minimum)5 of 7 positiveSources
PraiseData Retention and Security0 of 0 positive (under the minimum)1 of 1 positive (under the minimum)1 of 1 positive (under the minimum)Sources
PraiseTool Calling, Embeddings and Vision0 of 0 positive (under the minimum)1 of 1 positive (under the minimum)1 of 1 positive (under the minimum)Sources
PraiseHosted Model Answers0 of 0 positive (under the minimum)1 of 1 positive (under the minimum)1 of 1 positive (under the minimum)Sources
ConcernPrice Rises and Cached Tokens1 of 1 raise it (under the minimum)8 of 11 raise it9 of 12 raise itSources
ConcernCapacity and Rate Limits1 of 2 raise it (under the minimum)1 of 2 raise it (under the minimum)2 of 4 raise it (under the minimum)Sources
ConcernBilling Controls2 of 2 raise it (under the minimum)0 of 0 raise it (under the minimum)2 of 2 raise it (under the minimum)Sources
ConcernAPI Compatibility Gaps2 of 2 raise it (under the minimum)4 of 4 raise it (under the minimum)6 of 6 raise itSources
ConcernCatalog Gaps and Deprecations4 of 4 raise it (under the minimum)0 of 0 raise it (under the minimum)4 of 4 raise it (under the minimum)Sources
ConcernFine-Tuning and Adapters1 of 1 raise it (under the minimum)2 of 2 raise it (under the minimum)3 of 3 raise it (under the minimum)Sources
ConcernVaries by Model and Day2 of 3 raise it (under the minimum)2 of 2 raise it (under the minimum)4 of 5 raise itSources
DisagreementDoes Fireworks cost practitioners less than what they compared it with?Costs less 2, Costs more 0 (under the minimum)Costs less 5, Costs more 7Costs less 7, Costs more 7Sources
Current readEvidence mix45 sources60 sources76 sourcesSources

Fact checks

GILD's checks of claims against official documentation, kept apart from the source counts.

  • Low Per-Token Bills. Fireworks Serverless text and vision prices are input / cached input / output in USD per 1M tokens. Ember-1 Standard is US$3.00 / US$0.30 / US$15.00. Serverless pricing (text), checked Nov 1, 2026
  • Low Per-Token Bills. Fireworks embeddings are billed per 1M input tokens, at US$0.008 for base models up to 150M parameters. Serverless pricing (text), checked Nov 1, 2026
  • Low Per-Token Bills. Fireworks AI uses a usage-based pre-paid billing system: per token for serverless inference, per GPU usage time for on-demand deployments, and per training token for Managed Training. Cost FAQ, checked Nov 1, 2026
  • Low Per-Token Bills. The Fireworks Batch API processes large volumes of requests asynchronously at 50% off Serverless per-token prices. Not all models support the Batch API. Batch inference, checked Nov 1, 2026
  • Serving Speed. Fireworks deploys models on dedicated GPUs with fast autoscaling and minimal cold starts. Introduction, checked Nov 1, 2026
  • Serving Speed. On-demand deployments give you dedicated GPUs for your models. On-demand deployments, checked Nov 1, 2026
  • Serving Speed. Fireworks Serverless Inference is multi-tenant inference for popular open models running on Fireworks-managed infrastructure. Serverless overview, checked Nov 1, 2026
  • New Models Quickly. Fireworks lists 100+ supported models across text, vision, audio, image, and embeddings. Introduction, checked Nov 1, 2026
  • New Models Quickly. On-demand deployments let you upload your own models (for supported architectures) from Hugging Face or elsewhere. On-demand deployments, checked Nov 1, 2026
  • Moving Work to Open Models. You can use the OpenAI Python client library to interact with Fireworks. OpenAI compatibility, checked Nov 1, 2026
  • Moving Work to Open Models. Fireworks documents building production AI agents on Fireworks with LangChain, LlamaIndex, CrewAI, AutoGen, and other open-source frameworks using the OpenAI-compatible API. Agent frameworks, checked Nov 1, 2026
  • Moving Work to Open Models. Fireworks lists LiteLLM (native fireworks_ai provider) and Helicone (Fireworks proxy and provider integration) among the gateways that connect to your Fireworks account. LLM gateways, checked Nov 1, 2026
  • Moving Work to Open Models. Fireworks AI is a first-party inference provider inside Microsoft Foundry, with usage billed through Azure and counting toward your Microsoft Azure Consumption Commitment (MACC). Microsoft Foundry, checked Nov 1, 2026
  • Moving Work to Open Models. On Microsoft Foundry, Fireworks PayGo deployments run in the US Data Zone only. Foundry PTU (Provisioned Throughput) deployments provide dedicated GPU capacity reserved for your workload, with global region availability. Microsoft Foundry, checked Nov 1, 2026
  • Moving Work to Open Models. Fireworks' `auto` model ID picks a model for each new user turn. Closed models run on your own Anthropic or OpenAI account. To choose a model yourself, replace `auto` with its ID. When Fireworks' CLI connects Claude Code, Auto is the default model in `/model`. Nexus Auto, checked Nov 1, 2026
  • Moving Work to Open Models. Before you begin, create an API key in the Fireworks dashboard. Quickstart, checked Nov 1, 2026
  • Moving Work to Open Models. Fireworks is self-serve. Contact Fireworks for enterprise deployments. Pricing, checked Nov 1, 2026
  • Data Retention and Security. Fireworks has Zero Data Retention by default. Fireworks does not log or store prompt or generation data for any open models, without explicit user opt-in. For the Response API, with `store=True` (the default), stored conversation data is kept for up to 30 days. Setting `store=False` prevents any conversation data from being retained. Data handling, checked Nov 1, 2026
  • Data Retention and Security. Across every training surface, your training data is never used to train Fireworks-owned or shared models. Secure training, checked Nov 1, 2026
  • Data Retention and Security. By default, Fireworks supports Google SSO. If you have an enterprise account, Fireworks supports bringing your own identity provider using an OpenID Connect (OIDC) or SAML 2.0 provider. Coordinate with your Fireworks AI representative to enable the integration. SCIM provisioning is available for enterprise accounts and works with supported directory providers, including Okta, Microsoft Entra ID, and Google Workspace. Custom SSO must be configured before you can enable SCIM provisioning. To request SCIM enablement, contact your Fireworks AI representative, and Fireworks will enable Directory Sync for your account. API keys and service accounts are not blocked by SSO enforcement. SSO and SCIM, checked Nov 1, 2026
  • Data Retention and Security. Fireworks has achieved all three ISO certifications covering security, privacy, and AI management: ISO 27001, ISO 27701 and ISO 42001. Fireworks lists SOC 2 Type II as certified. Fireworks is HIPAA compliant. Data security, checked Nov 1, 2026
  • Data Retention and Security. Audit logs are available for Enterprise accounts. Audit logs include data access logs. All read, write, and delete operations on storage are logged, normalized, and enriched with account context. You can view audit logs with the Fireworks CLI or retrieve them with the List Audit Logs API. Audit logs, checked Nov 1, 2026
  • Data Retention and Security. On Enterprise accounts, data residency restricts inference on the account to a single region. Contact your Fireworks representative if you need this enabled. The listed restriction is US, on `us.api.fireworks.ai`. For a region that is not listed, contact Fireworks sales. Data residency, checked Nov 1, 2026
  • Tool Calling, Embeddings and Vision. Fireworks tool calling (also known as function calling) enables models to select and use external tools based on user input, through OpenAI-compatible tool specifications. Tool calling, checked Nov 1, 2026
  • Tool Calling, Embeddings and Vision. Fireworks hosts embedding and reranking models, useful for tasks like RAG and semantic search, and its embedding service is OpenAI compatible. Embeddings, checked Nov 1, 2026
  • Tool Calling, Embeddings and Vision. Fireworks vision-language models process both text and images in a single request, for uses including document analysis, chart interpretation and OCR, via serverless inference or dedicated deployments. Vision models, checked Nov 1, 2026
  • Capacity and Rate Limits. Fireworks says serverless users are subject to minor latency and performance variability during high volume periods. Serverless latency FAQ, checked Nov 1, 2026
  • Capacity and Rate Limits. When using Serverless, you may experience 429 Too Many Requests or 503 Service Overloaded. To avoid 429s, stay below the adaptive rate limits. To reduce the likelihood of 503s, upgrade to Priority tier. Serverless rate limits (text), checked Nov 1, 2026
  • Capacity and Rate Limits. All API usage on an account shares one request-throughput envelope: 10 RPM with no payment method or no credits, and 6,000 RPM (maximum) with a payment method and active credits. The 6,000 RPM cap applies account-wide as a fixed ceiling. Account quotas, checked Nov 1, 2026
  • Capacity and Rate Limits. The inference error codes page says there are no account-level rate limits on dedicated and on-demand deployments, and that a 429 there indicates the deployment's processing capacity is saturated. The account quotas page says on-demand and dedicated deployments are not limited by adaptive serverless TPM upper bounds, and that their requests still count toward account-wide request limits. Inference error codes, checked Nov 1, 2026
  • Capacity and Rate Limits. By default, Fireworks deployments are multi-region, and Fireworks can move and spread them across regions as needed. Supported multi-regions are GLOBAL, US, CANADA, EUROPE and APAC. New accounts receive GPU quota for the GLOBAL multi-region only. The US, CANADA, EUROPE and APAC multi-regions and all single regions start at zero quota and must be granted by Fireworks. The `--region` flag pins a deployment to a single region, and a deployment with a `--region` you have no quota for is rejected at creation. Regions, checked Nov 1, 2026
  • Billing Controls. If your balance reaches zero and Auto Reload is not enabled, usage pauses until you add credits. Billing and credits FAQ, checked Nov 1, 2026
  • Billing Controls. When usage reaches 100% of the monthly spend limit, all API requests pause automatically across serverless inference, deployments, and training. This suspension does not apply to Enterprise accounts. Account quotas, checked Nov 1, 2026
  • Billing Controls. Contracted customers may have the option to move to post-paid billing. Billing and credits FAQ, checked Nov 1, 2026
  • API Compatibility Gaps. Fireworks Anthropic compatibility serves only POST /v1/messages and does not support Anthropic server side tools. Anthropic compatibility, checked Nov 1, 2026
  • Fine-Tuning and Adapters. Fireworks offers supervised and reinforcement fine-tuning of models up to 1T+ parameters. Introduction, checked Nov 1, 2026
  • Fine-Tuning and Adapters. Fireworks serves fine-tuned models for the same price as base models. Pricing, checked Nov 1, 2026