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

Google published autofinetune, a demo in which an agent post-trains large language models overnight on Cloud TPUs using Tunix, Gemma, Antigravity CLI, and Gemini Flash 3.7.

Google showed an overnight loop in which an agent turns the knobs that fine-tune a language model. Treat that as Google’s demo, not proof that human engineers are obsolete.

Google’s Developers Blog post “Autonomous LLM post-training with Tunix on TPUs,” by Wei Wei, announces autofinetune. The idea is an autonomous research loop for post-training large language models (LLMs) — extra training after a model is already built. Two methods on the page: Supervised Fine-Tuning (SFT), which teaches from labeled examples, and reinforcement learning via Group Relative Policy Optimization (GRPO). The stack is Tunix, Gemma, and Cloud TPUs — Tensor Processing Units, Google’s training chips — orchestrated with Antigravity CLI and Gemini Flash 3.7. That company blog is the filing event. The setup details are in Sources.

On the post, a human writes a Markdown specification (program.md) that defines loop boundaries, evaluation criteria, and constraints, and provides execution code (run.py). An agent then 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.

First case study, SFT: FunctionGemma 270M — 270 million parameters, meaning 270 million tunable weights — on a mobile-actions task, run on a Cloud TPU v5e-1. Google reports about 20 automated experiments in a couple of hours. The objective was post-training evaluation accuracy on function-call generation. Google said the agent could tweak cheap add-on settings, not the dataset, the number of training passes, or the model shape.

Second case study, GRPO: Gemma 3 1B — 1 billion parameters — on GSM8K, a grade-school math word-problem set, via Tunix’s official GRPO example, on a Cloud TPU v6e-1. Google reports about 40 experiments over 2–3 days. The objective was combined numerical_accuracy plus format_accuracy. Google says the agent improved total reward by about 10% — one extra tenth on the score it was optimizing — via better add-on settings, sampling temperature, a KL (Kullback–Leibler) penalty, the 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, general-availability versus preview status, or production readiness beyond what this primary says. The post points to the autofinetune GitHub repository plus the Tunix library.