# Fireworks AI details speed gains for Kimi K3 reinforcement learning

> Fireworks AI reports up to 2.6× faster training and better legal‑task scores for its 2.8‑trillion‑parameter Kimi K3 model.

Oossa · 2026-10-23 · https://oossa.com/en/fireworks-ai-details-speed-gains-for-kimi-k3-reinforcement-learning

Fireworks AI released a technical report on how it speeds up reinforcement learning for its Kimi K3 model, a 2.8 trillion‑parameter system with a 1 million‑token context window. The company describes new kernels and communication tricks that give 1.3–2.6× faster forward and backward passes. Using the same setup, the Kimi‑based legal model Harvey Tenet scored 19.7% all‑pass on the LAB Contracts benchmark, beating the prior best Fable 5’s 11.5%.

## The facts

- Kimi K3 has 2.8 trillion parameters and a 1 M‑token context window.
- Harvey Tenet achieved 19.7% all‑pass on LAB Contracts, versus Fable 5’s 11.5%.

## Why it matters

The speed improvements let developers iterate on specialized AI policies faster, and the higher legal benchmark score shows the model can handle complex contract analysis better.

## Sources & references

1. [Kimi K3 on Fireworks: Fast RL at frontier scale](https://fireworks.ai/research/kimi-k3-on-fireworks-fast-rl-at-frontier-scale/) – Fireworks AI

Last updated: 2026-10-23