Fireworks: open models plus fine-tuning match closed ones — Cursor gets 13x faster inference
AI Engineer · youtube · 2026-10-12
A talk by Jetashree Ravi of Fireworks AI at AI Engineer World's Fair on moving from renting closed intelligence to a train-to-deploy loop.
Three truths: open models like GLM 5.2 are production-ready; customization closes the quality gap with closed models; the AI stack is shifting to a continuous train-and-serve loop.
Scale and results: 30 trillion tokens served, 10,000+ companies; Cursor's Composer 2 trained and hosted on Fireworks with 13x faster inference; 40x latency improvement for Vercel; 12% quality gain for Genspark via RFT.
Tech: FireAttention engine with custom CUDA kernels, speculative decoding with custom draft models, three training paths (training agent, managed training, training API for SFT/RFT/DPO), one-click deploys, monitoring, hot-loading weights with zero downtime, and numerics matching between training and serving.
More from Infra
- Decapybara: open-source C++20/CUDA rebuild of MegaCapybara inference server — giveen · 2026-10-12
- $2 ESP32 board runs Pi-hole-style DNS blocker with 140K domains in 0.7MB — LinusEkenstam · 2026-10-12
- Berth: open-source VRAM leasing daemon stops OOM crashes when multiple AI engines share one GPU — Distinct-Pie2389 · 2026-10-12
- Engineer joins NVIDIA's Groq LPU compilers team, working on multi-chip partitioning — blelbach · 2026-10-12
- Linus Ekenstam wants nothing less than a 100B-param model running on your phone — LinusEkenstam · 2026-10-12
- Google illegally cutting forest for data centers, Finland halts project over environmental review — heyimada · 2026-10-12