Fireworks AI

Fireworks AI is a generative AI infrastructure platform that helps companies turn open-source models into fast, cost-efficient, specialized intelligence they own end-to-end. Rather than locking developers into closed, proprietary APIs, Fireworks focuses on serving and training the best open-weight models, including DeepSeek, Kimi, GLM, Qwen, and gpt-oss, through a serverless inference API, on-demand GPU deployments, and a full spectrum of training options from guided, config-led runs to fully custom training loops on Fireworks' own infrastructure. The platform's proprietary optimizations, including the FireAttention CUDA kernel, speculative decoding, and the FireOptimizer adaptive serving engine, are built to deliver industry-leading throughput and latency at production scale.

Founded in late 2022 by CEO Lin Qiao and six co-founders who were core members of Meta's PyTorch team, Fireworks channels deep systems expertise in distributed training and inference directly into product performance. Its OpenAI and Anthropic-compatible serverless API lets teams switch from closed-model providers with minimal code changes, while on-demand and reserved GPU deployments give teams dedicated capacity across H100, H200, B200, B300, and GB300 hardware. Fireworks Nexus, a newer offering, acts as a drop-in replacement for closed-model coding APIs, intelligently routing requests to the best open or closed model per task and helping engineering teams cut AI coding spend by 50-75%.

Fireworks AI has scaled rapidly, processing tens of trillions of tokens daily for more than 10,000 enterprise customers, with Cursor, Vercel, Notion, Quora, Sourcegraph, and UiPath among its highest-profile users; Cursor uses Fireworks to serve and train the models behind its Composer coding agent at production scale, while Notion cut inference latency from roughly 2 seconds to 350 milliseconds after fine-tuning with the platform. Described by NVIDIA's Jensen Huang as "the TSMC of AI Factories," Fireworks has raised a $1.505 billion Series D at a $17.5 billion valuation as of 2026, crossing $1 billion in annualized revenue.

Fireworks AI's serverless inference is pay-per-token with postpaid billing and $1 in free starter credits, with per-model rates across Standard, Priority, and Fast tiers detailed in its documentation (e.g. GLM 5.2 at $1.40/M input and $4.40/M output tokens, MiniMax M3 at $0.30/M input and $1.20/M output tokens). Embeddings are priced by base model size, from $0.008 to $0.10 per 1M input tokens. Training is priced per 1M training tokens for supervised fine-tuning (SFT) and direct preference optimization (DPO): LoRA SFT ranges from $0.50 (models up to 16B parameters) to $10.00 (models over 300B), with Full Param SFT and DPO costing roughly 2-4x more depending on model size and method. Reinforcement fine-tuning is billed per GPU hour at on-demand rates. The Serverless Training API charges separately for prefill, cached prefill, sample, and train tokens (e.g. Qwen 3.5 9B at $0.66-$1.995 per 1M tokens depending on operation). On-demand GPU deployments are billed per GPU hour: $7.00 for H100 or H200, $10.00 for B200, $12.00 for B300, and $18.00 for GB300, with region-restricted (US/Europe) deployments priced at 1.5x standard rates. Reserved and enterprise capacity pricing is available by contacting sales.