Fireworks AI
by Fireworks AI
Own your specialized intelligence — inference, fine-tuning and RL for open models in one platform.
Fireworks AI is a production platform for running, tuning and training open-weight models. Teams replacing closed-model APIs get serverless per-token inference, dedicated GPU deployments and LoRA or full-parameter fine-tuning behind an OpenAI- and Anthropic-compatible interface, so they can own a specialised model rather than rent a general one.
Fireworks AI is a managed platform for running, tuning and training open-weight models in production. It offers three delivery modes: serverless pay-per-token inference, on-demand dedicated GPU deployments (H100 80GB and H200 141GB at $7.00/hour, B200 180GB at $10.00, B300 288GB at $12.00) with autoscaling, multi-region placement and minimal cold starts, and reserved capacity with higher quotas and early hardware access. All of it sits behind APIs that are deliberately OpenAI- and Anthropic-compatible, so migrating an existing application is close to a base-URL change. The catalogue spans more than 100 text, vision, audio, image and embedding models — DeepSeek v3.2 and v4, GLM 5.2, Kimi K3, Qwen variants, Gemma 4, MiniMax, FLUX.1 and Whisper V3 among them — with function calling, structured JSON output, reranking and batch inference treated as first-class rather than bolted on. Above inference sits a training stack: LoRA and full-parameter supervised fine-tuning plus DPO and reinforcement learning on models up to 1T+ parameters, priced per million training tokens from $0.50 (LoRA SFT under 16B) to $40 (full-parameter DPO above 300B), with checkpoints deploying straight to production. A newer product, Nexus, is pitched as a drop-in replacement for closed coding APIs claiming a 50–75% cut in AI coding spend. Founded in 2022 by former Meta PyTorch lead Lin Qiao and headquartered in Redwood City, California, Fireworks reached roughly $800M annualised revenue by May 2026 across 10,000+ customers including Cursor, Vercel, Notion, Sourcegraph, UiPath, Quora, Uber and Cresta, and raised a $1.505B Series D in July 2026 at a $17.5B post-money valuation led by Atreides Management, Index Ventures and TCV — after a $250M Series C at $4B in October 2025.
The platform or ML engineering team that has outgrown a closed-model API on cost and wants a tuned open model it controls — without standing up its own GPU fleet.
Fine-tune an open model on proprietary data and serve it in production from the same platform, with an OpenAI-compatible endpoint and SOC 2 Type II / HIPAA coverage.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Usage-based, Subscription, Contact for pricing
- Target Market
- CTOs, Enterprise Developers, ML Engineers, Data Scientists
- Deployment
- Cloud-first, API-based, Multi-cloud
- Founded
- 2022
- Headquarters
- Redwood City, California, United States
- Team Size
- 201-500
- Customers
- 10,000+ companies, including Cursor, Vercel, Notion, Sourcegraph, UiPath, Quora, Uber and Cresta
Key Features
- ✓Serverless per-token inference
Pay only for tokens consumed across 100+ open models, with Priority and Fast service options for latency-sensitive traffic.
- ✓On-demand dedicated GPUs
H100 and H200 at $7.00/hour up to B300 at $12.00/hour, with autoscaling, multi-region placement and minimal cold starts.
- ✓Managed fine-tuning and RL
LoRA, full-parameter SFT, DPO and reinforcement training on models up to 1T+ parameters, billed per training token.
- ✓OpenAI- and Anthropic-compatible APIs
Existing client code migrates with a base-URL and key change rather than an application rewrite.
- ✓Agentic primitives built in
Function calling, structured JSON outputs, reranking and embeddings ship as platform features, not third-party add-ons.
- ✓Nexus for coding workloads
A drop-in replacement for closed coding APIs that Fireworks claims cuts AI coding spend by 50 to 75 percent.
- ✓Batch inference
Asynchronous large-scale jobs for backlogs where unit cost matters far more than response latency.
Capabilities
Use Cases
- •Closed-model cost replacement
Swap a frontier closed API for a tuned open model to cut per-token spend while holding output quality steady.
- •Serving AI coding assistants
Cursor and Sourcegraph-class products push enormous token volume where both unit cost and latency shape retention.
- •Domain fine-tuning to production
Train LoRA or full-parameter adaptations on proprietary data, then deploy the checkpoint to dedicated GPUs immediately.
- •RAG and enterprise search
Hosted embedding and reranking models sit alongside generation, removing a separate vector-infrastructure vendor from the stack.
- •Regulated healthcare workloads
SOC 2 Type II and HIPAA compliance with zero prompt retention on open models makes PHI-adjacent deployments viable.
Ideal For
Best For
- ✓Replacing an expensive closed frontier API with a tuned open model to cut per-token cost at scale
- ✓AI coding assistants and copilots where token volume is enormous and latency shapes the product
- ✓Fine-tuning a domain model on proprietary data and deploying the checkpoint without changing platforms
- ✓RAG and search systems needing hosted embeddings, reranking and generation from one vendor
- ✓Regulated workloads that require SOC 2 Type II, HIPAA and zero prompt/generation retention by default
Not Ideal For
- ✗Organisations that must keep inference on their own hardware — Fireworks is cloud-only with no on-premises option
- ✗Small teams doing occasional experiments: there is no free tier, only $1 in starting credits, so evaluation costs money almost immediately
- ✗Teams whose fine-tuning volume is high enough that managed training markup outweighs renting raw GPUs — practitioners have publicly documented a $192 bill against roughly $0.17 of raw compute
- ✗Buyers who need transparent enterprise pricing up front, since reserved capacity and rate limits are entirely sales-negotiated
Integrations
Deployment
Market & Ratings
10,000+ companies, including Cursor, Vercel, Notion, Sourcegraph, UiPath, Quora, Uber and Cresta
Market Analysis
Pros
- ✓One platform covers serverless inference, dedicated GPUs, fine-tuning and reinforcement learning — no stitching vendors together
- ✓OpenAI and Anthropic API compatibility makes migrating off closed models near-trivial
- ✓Very broad open-model catalogue (100+ across text, vision, audio, image and embeddings), refreshed quickly after new releases
- ✓SOC 2 Type II, HIPAA and ISO certified with zero data retention on open models unless the customer opts in
- ✓Financial durability — roughly $800M annualised revenue by May 2026 and $1.8B+ raised — reduces vendor-failure risk
Cons
- ✗Managed training is priced far above raw GPU cost; a Hacker News post documented $192 charged against roughly $0.17 of compute
- ✗Only $1 in free credits and no free tier, so evaluation starts costing money almost immediately
- ✗Roughly 50% gross margins in a fast-commoditising market invite price pressure from self-hosted vLLM and SGLang
- ✗No on-premises option, which rules it out wherever data cannot leave the vendor's cloud
- ✗Enterprise pricing is opaque — reserved capacity, rate limits and volume discounts are entirely sales-negotiated
Pricing
Serverless
From $0.07 per 1M input tokens
- ✓Pay-per-token across 100+ models
- ✓$1 in free starting credits
- ✓Priority and Fast service options
- ✓Postpaid billing
On-Demand GPUs
From $7.00/hour
- ✓H100 80GB / H200 141GB at $7.00/hr
- ✓B200 180GB at $10.00/hr
- ✓B300 288GB at $12.00/hr
- ✓Autoscaling and multi-region deployment
Managed Training
From $0.50 per 1M training tokens
- ✓LoRA SFT from $0.50/1M tokens (models under 16B)
- ✓Full-parameter DPO up to $40/1M tokens (models over 300B)
- ✓Reinforcement learning supported
- ✓Checkpoints deploy straight to production
Enterprise / Reserved
Contact for pricing
- ✓Higher rate limits
- ✓Faster speeds and lower unit costs
- ✓Early hardware access
- ✓Reserved capacity
Three separate meters run in parallel: serverless tokens ($0.07 to $3.00 per million input depending on model), GPU-hours for dedicated deployments ($7 to $12 per hour), and training tokens ($0.50 to $40 per million by model size and method). There is no free tier — only $1 in starting credits — and enterprise rates, reserved capacity and higher rate limits are all quoted by sales rather than listed. Practitioners on Hacker News have publicly complained that managed training marks raw GPU cost up steeply, citing a $192 charge for roughly eight minutes of work.
Security & Compliance
Connect
Sources
This page was written from 5 sources, 3 on domains other than fireworks.ai.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
WEKA NeuralMesh
Microservices storage and memory fabric built for AI training and inference at exabyte scale
Tsuga
Bring-your-own-cloud observability that keeps telemetry, and its cost, inside your own AWS account
OpenObserve
Open-source, Rust-based observability on object storage — logs, metrics, traces and LLM telemetry in one binary
Coralogix
AI-native observability that queries logs, metrics and traces in place — index-free, in your own S3 bucket