Google shows autofinetune — agents that post-train LLMs overnight on TPUs

11 Sep 2026: Google published autofinetune — an autonomous research loop for LLM post-training (SFT and GRPO) using Tunix, Gemma, Cloud TPUs, Antigravity CLI, and Gemini Flash 3.7. Company Developers Blog is the primary. Case-study metrics stay Google’s.

11 Sep 2026: Google Developers Blog “Autonomous LLM post-training with Tunix on TPUs” (Wei Wei) announces autofinetune — applying autonomous research loops to LLM post-training (Supervised Fine-Tuning and Reinforcement Learning via GRPO) on Google’s stack: Tunix, Gemma, and Cloud TPUs, orchestrated with Antigravity CLI and Gemini Flash 3.7. That dated company blog is the filing event.

Paradigm on the post: a human writes a Markdown specification (program.md) defining loop boundaries, evaluation criteria, and constraints; provides execution code (run.py); an agent iterates — modify the script, run training, measure the metric, keep winning commits or revert regressions, and log results in results.tsv. Google says the loop is inspired by the earlier autoresearch pre-training framing on the same post.

Case study 1 (SFT): FunctionGemma 270M on mobile-actions; Cloud TPU v5e-1; about 20 automated experiments in a couple of hours; objective = post-training evaluation accuracy on function-call generation. Google’s stated boundaries: the agent may tune LoRA rank/alpha, target modules, learning rates, optimizers, batch size, and seeds — not the dataset, epochs, or architecture.

Case study 2 (GRPO): Gemma 3 1B on GSM8K via Tunix’s official GRPO example; Cloud TPU v6e-1; about 40 experiments over 2–3 days; objective = combined numerical_accuracy + format_accuracy. Google says the agent improved total reward by about 10% via better LoRA configs, rollout temperature, KL penalty, system prompt, and related knobs. Those figures are Google’s case-study claims — not independent replication.

This filing does not invent open-source license terms, pricing, GA versus preview status, or production readiness beyond what this primary says. Code pointer on the post: the autofinetune GitHub repository plus the Tunix library. This desk does not invent star or download counts.

A dated Google primary turns LLM post-training hyperparameter search into an agent loop on Tunix/TPU — desk keeps the overnight-agent framing as Google’s demo, not a proven replacement for human ML engineers.