Fireworks AI
first_party
Getting credentials
api-key
api_key via header (header Authorization: Bearer)
Environment: FIREWORKS_API_KEY
Create keys in the Fireworks console (app.fireworks.ai, Account Settings > API Keys).
Endpoints
| ID | URL | Protocol | Auth |
|---|---|---|---|
| inference-v1 | https://api.fireworks.ai/inference/v1/chat/completions | openai-chat | any |
| responses | https://api.fireworks.ai/inference/v1/responses | openai-responses | any |
| anthropic | https://api.fireworks.ai/inference/v1/messages | anthropic-messages | any |
API surfaces
textstreamingembeddingsbatchrerank
Quirks
- reasoning_effort (low/medium/high) AND an Anthropic-style thinking object {type: enabled, budget_tokens >= 1024} are both accepted on chat/completions — specifying both errors. (docs)
- Interleaved thinking: you MUST echo prior assistant reasoning_content when the last message is a tool result; reasoning_history: preserved retains reasoning across turns. (docs)
- Batch is 50% of serverless price via the Fireworks-native batch API (not OpenAI /v1/batches); Responses API stores by default; streaming includes usage in the final chunk. (docs)
Models offered
| Model | Wire ID | Endpoint | Status | Reasoning |
|---|---|---|---|---|
| deepseek/deepseek-v4-pro | deepseek-v4-pro | inference-v1 | ga | effort · default on · returns reasoning_content |
| moonshot/kimi-k2-6 | kimi-k2p6 | inference-v1 | ga | effort · default on · returns reasoning_content · round-trips reasoning_content |
| openai/gpt-oss-120b | gpt-oss-120b | inference-v1 | ga | effort · default on · mandatory · returns reasoning_content |
| zai/glm-4-7 | glm-4p7 | inference-v1 | ga | adaptive · default on · returns reasoning_content · round-trips reasoning_content |