Company

Fireworks AI

Fireworks AI is a US company that runs an inference and training platform for generative AI models, with a focus on open-weight models. Developers and companies use it to call models through a serverless API, rent dedicated GPU deployments, and fine-tune models on their own data, then serve the results from the same platform. Who it is for: teams building AI products and agents who want to use open models such as DeepSeek, Kimi, GLM or Meta's models without running their own GPU fleet, and enterprises that want more control over cost, data handling and model choice than a single closed-model API offers. Key products, as described by the company: - Serverless inference: pay-per-token access to hosted open models through an API, with higher serverless rate limits generally available since November 2026 and Elastic Burst, which lets usage exceed its normal ceiling when spare capacity exists. - Ultra-fast: a low-latency inference tier on NVIDIA Blackwell GPUs, in public preview since November 3, 2026. - Background Mode: a lower-cost runtime for work that can wait, which Fireworks describes as half price, alongside multi-day prompt caching for agents that pause and resume. - Training and serving of custom models: supervised fine-tuning, preference optimization and reinforcement learning, plus Specialized Model Inference for serving LoRA adapters serverlessly with no charge while idle (public preview). - Fireworks Nexus: model routing for individual developers inside tools such as the Codex app, Claude Desktop and OpenCode, via the Nexus CLI and Nexus Desktop, with Pro and Teams subscription plans. - Enterprise features: model governance, spending controls and exportable audit logs, plus Fireworks Sovereign (private preview), which runs the Fireworks inference engine in a customer's own cloud or data center. Pricing type: usage-based (per token or per GPU hour) for the platform, with paid subscription plans for Nexus and custom enterprise agreements. Check the official pricing page for current rates, since preview products may not yet list prices. Alternatives: other open-model inference providers such as Together AI and Groq, model gateways such as OpenRouter and Vercel AI Gateway, the model catalogs in Amazon Bedrock and Google Vertex AI, or self-hosting open weights on your own GPUs. Official site: https://fireworks.ai