Fireworks AI

Fireworks AI is a generative-AI infrastructure company founded in 2022 and headquartered in Redwood City, California. It was started by a team of seven engineers with backgrounds building and maintaining PyTorch and large-scale machine learning infrastructure at Meta and Google, including CEO Lin Qiao, formerly head of PyTorch at Meta. The company operates a training and inference platform that lets organizations deploy, fine-tune and serve open-source large language models -- including DeepSeek, GLM, Kimi, Qwen and Minimax -- through OpenAI- and Anthropic-compatible APIs. Its inference offering spans serverless pay-per-token access (priced from roughly $0.07 to $3 per million tokens depending on model and priority tier), on-demand dedicated deployments, and reserved capacity for high-volume workloads, running across multiple cloud regions on NVIDIA and AMD hardware. On the training side, Fireworks AI supports LoRA and full fine-tuning, guided training runs with cost estimation, and integrated reinforcement-learning pipelines for building custom or task-specific models. In August 2026, Fireworks AI began hosting Moonshot AI's Kimi K3 model on Microsoft Foundry, extending its role as a third-party inference backend embedded in a major cloud AI marketplace. Companies publicly named as customers on the Fireworks AI website include Cursor, Vercel, Notion, UiPath, Sourcegraph, Genspark and Motif. The company has raised venture funding from Benchmark, Sequoia Capital, Lightspeed Venture Partners and Index Ventures, with NVIDIA and AMD also participating as strategic infrastructure investors, and has been reported at a multi-billion-dollar valuation following funding rounds through 2025. According to a 2026 employer-profile snapshot, Fireworks AI has approximately 63 employees, a relatively lean headcount consistent with an infrastructure-focused engineering organization rather than a large services business. Fireworks AI is not a robotics hardware manufacturer; it is a horizontal AI model-serving and fine-tuning platform. Its relevance to the robotics and embodied-AI space comes from providing low-latency inference and custom-model training infrastructure that downstream developers -- including those building agentic software, multimodal systems and robotics-adjacent AI applications -- can use to deploy foundation models in production without operating their own GPU clusters.