fireworks.ai

Fireworks offers a full spectrum of training options, from guided runs to configuration-led training to custom training logic. Its inference engine is optimized for throughput and latency, with serverless, on-demand, and reserved deployment options. The platform provides access to a model library with the latest open-source models, including DeepSeek, GLM, Kimi, and Qwen, all optimized for cost, speed, and quality. It supports OpenAI and Anthropic compatible APIs, making integration easy.

Fireworks is used for AI coding assistants, such as Cursor's Composer 2, which relies on Fireworks for reinforcement learning inference. Vercel uses Fireworks to fine-tune models for its v0 product, achieving better than state-of-the-art performance. Notion reduced latency from 2 seconds to 350 milliseconds by fine-tuning with Fireworks. Other customers include Sourcegraph, Quora, and UiPath, who use Fireworks for hosting open models and powering AI features.

Fireworks offers a pay-per-token pricing model for serverless inference, with options for Priority and Fast tiers. On-demand and reserved deployments are also available for dedicated capacity. The model library shows specific pricing per million tokens, such as $0.14 for DeepSeek-V4-Flash and $3 for Kimi K3. Training costs are not detailed on the page, but users can request a demo or contact the team for more information.