Fireworks On-Demand Deployments
Run an open model on dedicated NVIDIA GPUs, from H100 to GB300, billed per GPU-second.
Overview
On-Demand Deployments give a model its own dedicated GPUs on Fireworks, for steady, high-volume or latency-sensitive production use. You choose the model and GPU type, and pay per GPU-second with no extra charge for start-up time. Per-GPU prices: NVIDIA H100 80 GB and H200 141 GB at US$8 an hour, B200 180 GB at US$13, B300 288 GB at US$15 and GB300 288 GB at US$20. Deployments locked to a specific region cost 1.5 times as much. For larger commitments, Fireworks offers enterprise reserved capacity through its sales team. Deployments use the same OpenAI-compatible API as Serverless, and can run your own fine-tuned models. Prices change often, so check Fireworks' pricing page for current rates.
Specifications
| Hosting type | Dedicated endpoint |
|---|---|
| Model families | Llama, Mistral, DeepSeek, Qwen, Gemma, gpt-oss |
| LLM models | Llama 3.3 70B Instruct, Mistral NeMo, DeepSeek-R1, DeepSeek-V3.1, Qwen2.5 72B Instruct, Qwen2.5-Coder 32B Instruct, Qwen3 32B, Qwen3 30B-A3B, Qwen3 235B-A22B, Gemma 2 9B, Gemma 3 4B, Gemma 3 12B, Gemma 3 27B, gpt-oss-120b, gpt-oss-20b |
| Pricing range | US$8 – US$20 |
| Pricing model | Per GPU-hour, Reserved or committed capacity |
| Free tier available | No |
| Data center locations | |
| Hardware / GPU options | NVIDIA H100 80 GB, H200 141 GB, B200 180 GB, B300 288 GB and GB300 288 GB GPUs. |
| VRAM & memory | 80 GB – 288 GB |
| Maximum model size | 671B |
| OpenAI-compatible API | Yes |
News about Fireworks AI
All news →Fireworks AI: serverless and on-demand hosting now in our directory
We've added Fireworks AI, the fast open-model inference platform founded by Meta's former PyTorch leads, with its serverless and on-demand hosting plans and 17 supported models.
