Coding & Development

Fireworks AI

Inference and fine-tuning platform for open-source AI models

Serverless pay-per-token (prepaid credits; per-model rates in docs); on-demand GPUs from $8.00/hr (H100/H200); fine-tuning from $0.50/1M training tokens; Enterprise via sales. Free credits not confirmed on the pricing page.
Visit Fireworks AI →
Pricing
Serverless pay-per-token (prepaid credits; per-model rates in docs); on-demand GPUs from $8.00/hr (H100/H200); fine-tuning from $0.50/1M training tokens; Enterprise via sales. Free credits not confirmed on the pricing page.
Best for
Who it is for: developers and ML teams shipping production apps on open-weight models who need low latency, fine-tuning or dedicated capacity.
Official site
fireworks.ai
Last checked
2026-11-11

Fireworks AI is a good fit for engineering teams that want to run open-weight models in production through an API, and optionally fine-tune them, without running their own GPUs. Skip it if you want a free sandbox, a no-code tool, or mainly need the newest closed frontier models. The company was founded in 2022 by former Meta PyTorch engineers, led by Lin Qiao. It offers serverless pay-per-token inference, dedicated on-demand and reserved GPU deployments, and managed training from guided to fully custom. Serverless is compatible with OpenAI and Anthropic style APIs, so moving an existing app over is usually a small code change. Its site also promotes Nexus, a router that sends each task to an open or closed model and claims to cut AI coding spend by 50 to 75% (a vendor claim we have not tested).

Best for

Who it is for: developers and ML teams shipping production apps on open-weight models who need low latency, fine-tuning or dedicated capacity. Who it is NOT for: hobbyists wanting free usage, non-technical users, teams that only use closed models from a single lab, or anyone who wants to run models fully on their own machine (see Ollama).

Key features

01

Serverless inference

Pay-per-token API for hosted open models, billed on input, cached and output tokens, with Standard, Priority and Fast serving paths described in the docs.

02

On-demand deployments

Dedicated GPUs billed per GPU-second with no extra charge for start-up time; region-restricted deployments carry a 1.5x premium.

03

Reserved capacity

Guaranteed capacity for steady, high-volume workloads, priced through sales.

04

Managed fine-tuning

LoRA and full-parameter SFT and DPO, priced per million training tokens by base-model size; tuned models are served at base-model prices.

05

Training APIs

A serverless Training API with per-token prefill, sample and train rates on selected models, plus a dedicated option priced at GPU-hour rates.

06

Nexus router

Drop-in replacement for closed-model APIs that routes each task to an open or closed model.

Pricing breakdown

Serverless Inference

Pay per token
Prepaid credits with auto-reload and spend limits
  • Billed on input, cached and output tokens
  • Rates vary by model and serving path (Standard, Priority, Fast); see docs.fireworks.ai/serverless/pricing
  • Free credits not listed on the pricing page

Embeddings

From $0.008/1M input tokens
Pay per token (limited-time offer per page)
  • Up to 150M params: $0.008
  • 150M-350M: $0.016
  • Qwen3 8B: $0.10

Fine-Tuning

From $0.50/1M training tokens
Per 1M training tokens, by base-model size
  • LoRA SFT $0.50 and LoRA DPO $1.00 (up to 16B)
  • Full-parameter SFT $1.00 and DPO $2.00 (up to 16B)
  • Up to $10-$40 per 1M for models over 300B
  • Tuned models served at base-model price

On-Demand Deployments

From $8.00/hr (H100 or H200)
Per GPU-second
  • H100 80GB and H200 141GB: $8.00/hr
  • B200: $13.00/hr
  • B300: $15.00/hr; GB300: $20.00/hr
  • No extra charge for start-up time; 1.5x for region-restricted

Enterprise

Custom
Contact sales
  • Faster speeds, lower costs and higher rate limits
  • No published prices

Pros and cons

Pros

  • No charge for GPU start-up time and per-second billing on dedicated GPUs
  • Fine-tuned models cost the same to serve as the base model, which keeps customization affordable
  • One vendor for serverless inference, dedicated GPUs, reserved capacity and fine-tuning
  • OpenAI- and Anthropic-compatible APIs make migrating an existing app easier
  • Publicly named customers include Notion, Quora, Cursor and Vercel (testimonial speedups are vendor-reported)

Cons

  • Per-token serverless rates are not shown on the pricing page, so budgeting needs a trip to the docs and varies by model and serving path
  • No free credits are listed on the pricing page; earlier versions of this page cited $1 starter credit, which we could not confirm
  • On-demand GPU rates are high ($8 to $20 per hour) and only make sense at steady volume
  • Some embeddings pricing is labelled limited-time, so it may change
  • Enterprise pricing is sales-only

Alternatives to Fireworks AI

Frequently asked questions

Does Fireworks AI have a free tier?

The pricing page lists no free tier or starter credits as of 2026-10-20. It shows a time-limited free promotion for one training model (Muse Glimmer 30B SFT, until 11/30). Check the signup flow for current offers.

How is serverless inference priced?

Per token, with separate input, cached-input and output rates that depend on the model and on the Standard, Priority or Fast serving path. Exact rates are in the docs, not on the main pricing page.

How much do dedicated GPUs cost?

H100 and H200 are $8.00/hr, B200 $13.00, B300 $15.00 and GB300 $20.00, billed per GPU-second. Region-restricted deployments cost 1.5x.

Can I fine-tune models?

Yes. LoRA and full-parameter SFT and DPO are priced per 1M training tokens by model size, starting at $0.50, and tuned models are served at the base model's price.

Is it a good choice if I want to run models locally?

No. Fireworks is a hosted service. For local use, a tool like Ollama is the better match.

Ready to try Fireworks AI?

Head to the official site to explore pricing and start a free trial where available.

Visit Fireworks AI →