Tunix

Tunix is Google’s open-source library for post-training large language models, built on JAX and tuned for state-of-the-art performance on TPUs. It has around 2,500 GitHub stars and is under very active development, with commits landing in October 2026.

Overview

Short for “Tune-in-JAX”, Tunix streamlines everything that happens after pre-training: supervised fine-tuning, reinforcement learning and agentic RL, in one lightweight codebase. Where many post-training stacks are PyTorch-only, Tunix is aimed squarely at the JAX/TPU world, with sharding and performance work handled by the same organisation behind JAX itself. It integrates with Flax NNX for modelling and with high-performance inference engines such as vLLM and SGLang-JAX to generate rollouts during RL training. Google presents it as the post-training companion for its open models such as Gemma, and its documentation covers full fine-tuning as well as parameter-efficient approaches. A V2 release is out and the repository sees fresh commits in 2026 — this is a living Google project, not a research dump.

Key Features

• Supervised fine-tuning for LLMs at TPU scale
• Reinforcement learning, including agentic RL recipes
• Integrates Flax NNX modelling with vLLM and SGLang-JAX rollouts
• Designed for TPU performance with JAX throughout
• Parameter-efficient tuning options alongside full fine-tuning
• Example recipes and notebooks for common post-training jobs

💳 Pricing

Tunix is free and open source under the Apache-2.0 license, and there is no commercial tier or hosted upsell. Your costs are the TPU or GPU compute you train on.

Who Is It For

ML engineers post-training open models — especially Gemma — on Google Cloud TPUs, JAX teams who want SFT and RL in one library, and researchers reproducing published RL results.