google/tunix

Health v2 · maintenance only

Adoption not part of the score

2490 stars · 372 forks observed · 2026-10-21

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Tunix is a lightweight JAX-based library for post-training large language models, supporting supervised fine-tuning, preference optimization (DPO/ORPO), and reinforcement learning methods like PPO and GRPO. It is optimized for TPU/GPU performance and integrates with Flax NNX and inference engines like vLLM and SGLang-JAX.

Use cases

• fine-tune an LLM with LoRA or full weights
• run GRPO reinforcement learning post-training on TPUs
• train a model with DPO preference optimization
• do agentic RL with multi-turn tool use
• post-train LLMs in JAX with vLLM rollouts

When to choose

• you train or fine-tune LLMs in the JAX/TPU ecosystem
• you need SOTA RL post-training algorithms like GRPO or DAPO
• you want efficient TPU performance with Flax, MaxText, and vLLM integration

When to avoid

• you prefer PyTorch-based training stacks like TRL or Axolotl
• you need simple inference only without training
• you need a fully stable, feature-frozen library since Tunix is under active development

Facets

library · maturity active

llm-training machine-learning reinforcement-learning gpu-computing large-language-models machine-learning deep-learning reinforcement-learning python cloud jax post-training fine-tuning sft dpo grpo ppo tpu flax agentic-rl peft lora ai-agents gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="google/tunix")