Overview
Fireworks AI is an inference platform built for teams shipping open-weight models to production rather than experimenting in a notebook. It was started by former Meta PyTorch and Google engineers and focuses on speed: its FireAttention engine and adaptive speculative decoding deliver some of the highest tokens-per-second in the industry for models like Llama, DeepSeek, Qwen, and GLM. The catalog spans 200+ models across text, vision, image, embedding, and speech modalities, and new open releases typically land within 24 hours. What separates Fireworks from a bare token API is the full post-training stack under one key — you can fine-tune with LoRA or full SFT, host hundreds of adapters, and run batch jobs at half price, all without standing up your own vLLM cluster. The API is OpenAI-compatible, so most apps migrate by swapping the endpoint and key. It is SOC 2, HIPAA, and GDPR compliant with zero data retention options, which matters for enterprise workloads.
Key Features
- Serverless inference from $0.10/MTok — pay per token, no cold boots, no GPU provisioning for models under 4B params
- Fine-tuning + LoRA hosting — self-serve SFT, LoRA, and RL fine-tuning; serve many adapters in production from one API
- Batch API at 50% off — cheaper offline processing for document jobs, embeddings, and labeling
- Dedicated GPUs on demand — A100/H100/H200/B200 with auto-scaling to zero and predictable capacity
- Day-0 model support — new open weights go live within 24 hours of release
- OpenAI-compatible API — migrate from OpenAI by changing the endpoint and key
Pricing
| Plan | Price | For |
|---|---|---|
| Free starter | $1 credit | Initial testing across 50+ models |
| Serverless | $0.10–0.90/MTok | Pay-per-token dev and production (by model size) |
| Batch API | 50% of serverless | Bulk offline transcription, embeddings, labeling |
| Dedicated GPU | $2.90–9/hr (A100–B200) | Guaranteed capacity, always-on workloads |
| Enterprise | Custom | SLAs, BYOC, dedicated support |
Comparison
vs. Replicate: Replicate is simpler for spinning up community models by version, while Fireworks is faster and adds fine-tuning plus batch discounts under one API. vs. Together AI: Both are full-stack inference platforms; Together has a larger catalog, but Fireworks tends to win on raw speed and the 50% batch discount. vs. fal.ai: fal.ai specializes in media (image/video) inference with low latency; Fireworks covers broader modalities and adds post-training.