Fireworks AI

Model

1. SupportsModel APIs: Supports

   Model APIs, supports. Fireworks Serverless Inference is multi-tenant inference for popular open models, and the OpenAI Python client library works against it. The models themselves come from their makers, and closed models that Fireworks' auto model ID picks run on your own Anthropic or OpenAI account.

2. Model gateway: Supports

   Model gateway, supports. Fireworks' auto model ID picks a model for each new user turn. LiteLLM and Helicone connect to a Fireworks account, and routing across several hosts stays with a gateway like them.

Build

1. TouchesOrchestration: Touches

   Orchestration, touches. Fireworks documents building agents with LangChain, LlamaIndex, CrewAI, AutoGen and other open-source frameworks over its OpenAI-compatible API. The framework controls the agent's steps.

2. Tool use: Touches

   Tool use, touches. Tool calling lets hosted models select and use external tools through OpenAI-compatible tool specifications, and the tools run in your application. On the Anthropic-compatible endpoint, server-side tool execution is not supported.

Data

1. Retrieval: Supports

   Retrieval, supports. Fireworks hosts embedding and reranking models for RAG and semantic search through an OpenAI-compatible embedding service. The index and search stay with a vector database.

Evaluate

1. Evaluation: Touches

   Evaluation, touches. Supervised and reinforcement fine-tuning adapt models up to 1T+ parameters on Fireworks, and fine-tuned models serve at base-model prices. Test sets and scoring stay with an evaluation tool.

Operate

1. Observability: Touches

   Observability, touches. Helicone connects as a Fireworks proxy and provider integration, and Enterprise accounts get audit logs of storage reads, writes and deletes. Request tracing stays with an observability tool.

2. Fireworks AICoreCloud deployment: Core

   Cloud deployment, core. This is Fireworks' primary role: Serverless shares a pool of GPUs with models pre-provisioned, and on-demand deployments reserve dedicated GPUs with fast autoscaling. New accounts get GPU quota in the GLOBAL multi-region only, and other regions need quota granted by Fireworks.

3. Guardrails: Touches

   Guardrails, touches. Content moderation is one of the listed uses of the vision-language models Fireworks hosts. Screening prompts and outputs for attacks stays with a guardrail product.