Fireworks AI

Fireworks AI is a Redwood City, California inference platform, co-founded in 2022 by CEO Lin Qiao, who previously led the PyTorch team at Meta, with a founding team drawn largely from that group. Rather than training its own foundation models, Fireworks specializes in serving proprietary and open-source models through its own optimized stack, and it counts DeepSeek, Kimi, GLM, and Qwen among the open-weight families it hosts. On September 23, 2026 it shipped Ember-1, the first model in a new Fireworks Research series (the company had post-trained open weights before, including the Llama 3-based FireFunction-v2): a reasoning model post-trained from Moonshot's Kimi K3 that Fireworks says produces roughly 40 percent fewer reasoning tokens than K3 at comparable quality. Released as a Research Preview on Fireworks Serverless at $3 per million input tokens and $15 output, the same rate card as Kimi K3 itself, Ember-1 scores 92.2 percent on SWE-bench Verified against 93.2 for K3 at max settings on Fireworks' own table, essentially holding capability while cutting token spend. All figures are vendor-reported and no independent evaluator has scored Ember-1 yet.