AI Coding Live News

Fireworks said partner LangChain generates billions of agent-trace tokens a day, making frontier closed judges too costly. Fine-tuning a Qwen base on Fireworks matched GPT-5.5 / Opus-level judging at up to about 100x lower cost, and Fireworks is pushing its Training API for similar workflows.

Fireworks announced general availability of its Training API and Fireworks Lab, aiming to make model specialization accessible beyond labs that already own training clusters. The API lets teams connect a custom Python training loop to Fireworks-managed distributed training and rollout infrastructure, keeping control of loss, reward, data, and environment. Compute is split into Serverless mode for fast experiments and Dedicated mode when scale, control, and GPU economics matter. The stack is built for reinforcement learning as a single train-sample-serve-retrain loop, with three hard requirements: rollout throughput, weight sync so fresh weights land on the rollout deployment every step, and correctness so trainer and inference engine agree on the same model. Fireworks Lab additionally embeds forward-deployed researchers and engineers to co-design or build a model. Named customers include Cursor, Harvey, Figma, Vercel, Cognition, Factory, and Heidi.

Fireworks spotlighted Factory's Droid Shield as a Training API customer story. Factory's droids write and commit code faster than human reviewers can keep up, so secret detection has to catch real leaks without flooding the team with false alarms. Using the Fireworks Training API, Factory fine-tuned an open Qwen model that caught almost 20 percent more real secrets than GPT-5.5, at lower cost and latency. The result is a specialized scanner sitting in the commit path rather than a generic frontier chat model asked to grep for keys. For coding-agent vendors, it is a concrete example that a domain-fine-tuned open-weight model can beat a larger closed model on a verifiable production task, then ship at a cost that matches high-frequency agent traffic. It also shows why training and inference on one platform matters: the specialized weights have to stay cheap enough to run on every commit.

Fireworks AI detailed Harvey's proprietary model Tenet, post-trained from Kimi K3 using async RL on Fireworks Training API, doubling LAB Contract all-pass rates from 10.8% to 19.7% at matched cost.