Google introduces Tunix, an open-source JAX-native library designed for LLM post-training and model alignment. The library provides modular APIs for common workflows including PEFT training, DPO, PPO, GRPO, and distillation, optimized for TPU performance. Initial benchmarks show a 12% relative improvement in GSM8K math
Table of contents
What is available in this initial releaseQuantitative resultsTrusted by researchers and innovatorsCommunity and Collaboration - Get InvolvedSort: