# Fireworks AI

Canonical: https://inetgeek.com/llm-apis/fireworks/

Serverless inference for open-weight models from other labs, priced per token, with fine-tuning and dedicated GPUs billed separately.

## Who it suits

Running open-weight models without operating GPUs — the same DeepSeek and Qwen checkpoints the labs publish, served per token, with a path to dedicated hardware if throughput demands it.

Consider something else if You want a frontier proprietary model: Fireworks serves other people's open weights, so the ceiling is whatever the open ecosystem has released.

## What the documentation says

### Pricing

- Free tier: Yes ([source](https://fireworks.ai/pricing), read 2026-09-06)
- Pricing model: Three products on one bill. Serverless inference is per token, per model. Fine-tuning is per million TRAINING tokens by parameter band — $0.50 for models up to 16B, $10.00 above 300B. Dedicated GPUs are hourly: $7.00 for an H100 or H200, $10.00 for a B200. ([source](https://fireworks.ai/pricing), read 2026-09-06)

### Inference

- Input price, top model: $3.00 per million input tokens for Kimi K3, the most expensive model in its serverless catalogue. Fireworks hosts other labs' open-weight models rather than its own, so "flagship" here means the top of what it serves. ([source](https://fireworks.ai/models), read 2026-09-06)
- Output price, top model: $15.00 per million output tokens for Kimi K3. ([source](https://fireworks.ai/models), read 2026-09-06)
- Input price, cheapest model: $0.15 per million input tokens for GLM 5.3 Flash, with output at $0.50. DeepSeek-V4-Flash is $0.22 in and $0.66 out — the same model DeepSeek itself sells at $0.44 peak. ([source](https://fireworks.ai/models), read 2026-09-06)
- Context window: 1,048,576 tokens on Kimi K3, GLM 5.3 and both DeepSeek V4 variants; 262,144 on Qwen3.8 and 131,072 on Muse Glimmer 30B. The window is the model's, not Fireworks'. ([source](https://fireworks.ai/models), read 2026-09-06)
- Max output tokens: No separate output cap. max_tokens is bounded by the model's context length — prompt plus completion must fit inside it — and what happens on overflow is set by context_length_exceeded_behavior rather than a fixed ceiling. ([source](https://fireworks.ai/docs/api-reference/post-chatcompletions), read 2026-09-07)
- Prompt caching: Supported ([source](https://fireworks.ai/pricing), read 2026-09-06)

### GPU

- H100 SXM, per GPU-hour: $7.00 per hour for an H100 80GB on demand, or $8.00 on the shorter commitment shown beside it. The same rate applies to an H200 141GB. ([source](https://fireworks.ai/pricing), read 2026-09-06)
- Largest GPU offered: 180GB per GPU on a B200, at $10.00 an hour on demand. ([source](https://fireworks.ai/pricing), read 2026-09-06)

## Compared with

- [Fireworks AI vs DeepSeek](https://inetgeek.com/compare/deepseek-vs-fireworks/)
- [Fireworks AI vs Google Gemini API](https://inetgeek.com/compare/fireworks-vs-google-gemini/)
- [Fireworks AI vs Anthropic](https://inetgeek.com/compare/anthropic-vs-fireworks/)
- [Fireworks AI vs OpenAI](https://inetgeek.com/compare/openai-vs-fireworks/)
- [Fireworks AI vs DeepInfra](https://inetgeek.com/compare/deepinfra-vs-fireworks/)
- [Fireworks AI vs Baseten](https://inetgeek.com/compare/baseten-vs-fireworks/)
- [Fireworks AI vs Novita AI](https://inetgeek.com/compare/fireworks-vs-novita/)