Tunix
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 2
- Tutorial: Train a multi-turn tool-using agent with Tunix
Google Developers Blog published a tutorial for training a multi-turn, tool-using LLM agent with Tunix, Gemma 4 E2B, and Cloud TPU v6e-8. The example uses the Plane Strike game and reports that 100 steps of agentic GRPO raised held-out win rate from 9.4% to about 95% while reducing average game length to roughly 16 strikes. The complete training code is available in a GitHub repository.
Google Developers Blog · 🔥 0 - Google's Tunix Zero-TIM targets bitwise-zero train-inference mismatch in TPU RL
Google announced an implementation of zero train-inference mismatch (Zero-TIM) developed in Tunix, its open-source JAX-native library for LLM reinforcement learning and post-training on Cloud TPU. The approach makes the sampler and trainer assign bitwise-identical logprobs to the same action and token history, verified at multi-host scale across 64 TPU v5p chips with a DP8×TP8 mesh, by locking kernel accumulation order, ordering cross-shard reductions, sharing the action-logprob path and differentiating serving kernels with custom rules. In single-seed FrozenLake runs with a public Qwen3-8B checkpoint, Zero-TIM logged zero mean absolute sampler-trainer logprob difference at every step and reached higher observed solve rates than standard training with trainer-old logprobs and token-level truncated importance sampling.
Google Developers Blog · 🔥 0
Experience and discussion from the community
Share my Tunix experienceAsk about Tunix
Nobody has shared their experience with Tunix yet.