Fireworks AI
4.4
Fireworks AI offers serverless inference, fine-tuning and dedicated GPU deployments for open models. Serverless is billed per token against prepaid credits; supervised fine-tuning with LoRA runs $0.50 to $10.00 per million training tokens depending on model size, with DPO roughly double. On-demand H100 80GB is $8.00 an hour and B200 $13.00, with a 1.5x premium for region-restricted deployments.
- Commercial use OK
- Enterprise contract
- Pricing
- 従量課金
- Free tier
- No
- Commercial use
- Permitted
- Japanese UI
- Not confirmed
- Enterprise contract
- Available
What it is good for
- Low-latency serving of open models
- Fine-tuning with LoRA at published rates
- Dedicated GPUs rented by the hour rather than committed
- Teams needing region-restricted deployment
Strengths and trade-offs
Strengths
- Fine-tuning rates are published per million tokens, which is rare
- Region restriction is available as an explicit option
- Focused on inference speed
Trade-offs
- No general free tier
- H100 at $8.00 an hour is above some competitors
- Main model token rates live in separate documentation
Editors' review
Fine-tuning pricing you can actually budget
Most providers make you contact sales to find out what fine-tuning costs. Publishing a per-million-token range lets you estimate before committing engineering time, which changes whether the experiment happens at all.
The region-restriction premium is worth noting early if data residency is a requirement — it is a 50% uplift, not a rounding error.
Verdict
Good for fine-tuning experiments. Compare GPU-hour rates against Together before committing to dedicated.
Compiled from official documentation and public sources, with our own editorial judgement. We do not publish figures we cannot source.