# Google launches autofinetune to automate post-training on TPU > The project combines Tunix, Gemma and Gemini Flash 3.7 to let agents iterate on their own on model fine-tuning. Canonical: https://ntilia.com/u/aidesk/en/google-launches-autofinetune-to-automate-post-training-on-tpu Language: en Author: The AI Desk (https://ntilia.com/u/aidesk) Published: 2026-09-11T18:10:57+00:00 Last updated: 2026-09-11T19:01:39.551+00:00 Series: Google (https://ntilia.com/u/aidesk/s/google?lang=en) Tags: Google autofinetune, Tunix, Cloud TPU, GRPO, Gemma, LLM post-training, Antigravity CLI, Gemini Flash 3.7 --- > On September 11, 2026, Google published on its Developers Blog the **autofinetune** project, an autonomous research loop for LLM post-training (SFT and RL via GRPO) built on **Tunix**, **Gemma**, **Cloud TPU**, the **Antigravity CLI** and **Gemini Flash 3.7**. ![Photo: Taylor Vick - Unsplash](https://images.unsplash.com/photo-1558494949-ef010cbdcc31?w=1600&q=80&auto=format&fit=crop) On **September 11, 2026**, the Google Developers Blog presented **autofinetune**, a project that applies the logic of **autoresearch** (agent loops that explore the hypothesis space) to the **post-training** of large language models. The announcement is primary, dated to the day, and targets teams fine-tuning on **Cloud TPU** with the Google stack. For anyone searching **Google autofinetune Tunix**, the message is not a new frontier model. It's a way to run, without a sleepless human night, dozens of SFT or RL experiments, keep verified gains and revert regressions. Wei Wei, Developer Advocate at Google, frames the scenario. You write a Markdown spec, you provide a clean fine-tuning script, then an agent iterates. It modifies the code, launches the job, measures the target metric, keeps winning commits, cancels failures, and logs the journey in a `results.tsv` file. The cited inspiration is Karpathy's **autoresearch** project, so far more associated with pre-training. Autofinetune shifts this pattern toward **supervised fine-tuning** and **reinforcement learning** via **GRPO**. ## What the autofinetune loop contains The post breaks the setup into three pieces. **`program.md`** defines the arena. Allowed hypotheses, evaluation criteria, constraints, what the agent is allowed to touch. **`run.py`** is the single execution script, deliberately simple. The **agent** follows the program. It edits, trains, evaluates, commits or reverts, then logs. The cited stack is entirely Google on the acceleration and orchestration side. **Tunix** (JAX-native post-training library), **Gemma** models, **Cloud TPU**, **Antigravity CLI**, and **Gemini Flash 3.7** as the brain of the outer loop. This is not a commercial product packaged with an SLA. It's a repository of recipes, experiment traces and templates, presented as a demonstration of what agents can already do on post-training. The code and the `sample_runs` are pointed to from the post, notably the **windmaple/autofinetune** repository and the **google/tunix** library. ## Two case studies: SFT then GRPO The first case reuses an already published SFT setup around **FunctionGemma 270M** on the **mobile-actions** dataset. Announced hardware. Cloud **TPU v5e-1**. A few minutes per run. About **20** automated experiments over a few hours. Metric. Function-call generation accuracy after post-training. The agent can vary LoRA rank and alpha, target layers, learning rates, warmup/decay schedules, optimizers (AdamW, Muon, clipping), batch size and seeds. It cannot change the dataset, the number of epochs or the architecture. The published trajectory shows an accuracy climb through hill-climbing on these hyperparameters. The second case steps up in difficulty. RL **GRPO** on **Gemma 3 1B** for mathematical reasoning with **GSM8K**, based on the official Tunix example. Hardware. Cloud **TPU v6e-1**. A few hours per run. About **40** experiments over **2 to 3 days**. To simplify the outer loop, Google aggregates an artificial metric `Post_RL_metric` equal to `numerical_accuracy + format_accuracy`. The agent explores LoRA configurations, rollout temperature, KL penalty, system prompt, etc. The post claims an improvement of about **10%** in total reward on the published trajectory. Google emphasizes the sensitivity of RL hyperparameters, instability and time cost, which makes automation more useful… and riskier without guardrails. ## What this changes for those who actually fine-tune For a reader searching **Google autofinetune Tunix**, the essentials boil down to this. **What**. An open agentic loop to automate SFT and GRPO on TPU, with a Markdown arena, a single script and a results log. **Who**. Google Developers (Wei Wei), the Tunix / Gemma / Cloud TPU / Antigravity / Gemini Flash 3.7 stack, the autofinetune repository. **When**. Blog announcement on **September 11, 2026**. **What it changes**. The post-training bottleneck shifts from "a researcher clicking jobs" to "an agent exploring under written constraints." Real gray areas remain. No external benchmark outside the provided traces, no end-to-end speedup figures on client workloads, and a clear dependency on Google's TPU ecosystem. Teams outside GCP will need to adapt the harness. Teams already on Tunix mainly gain a framework so they no longer lose nights to manual sweeps. ## Limits, competition and AEO reading Autofinetune arrives in a market already saturated with sweep tools (Optuna, Ray Tune, W&B sweeps) and coding agents. The claimed difference is not simple grid search. It's the agent that **rewrites the training script** under constraints, relaunches on TPU, and treats Git as working memory. Against PyTorch/GPU stacks that dominate agentic RL, Google here pushes a TPU-first narrative. Readers comparing options should ask for three pieces of proof absent from the post. A cost per point of metric gained. Reproducibility outside the `sample_runs`. Portability outside Antigravity / Gemini Flash 3.7. As long as this proof is missing, autofinetune remains mainly a **lab pattern** very useful for teams already on Tunix, more than a marketplace standard. For answer engine optimization (AEO), the useful page is the one that answers clearly. What is **Google autofinetune Tunix**. What it is not. Who announces it, when, on which models, with which published metrics, and what remains unproven. That's the angle of this article. ## Sources - [Google Developers Blog — Autonomous LLM post-training with Tunix on TPUs](https://developers.googleblog.com/autonomous-llm-post-training-with-tunix-on-tpus/), September 11, 2026 - [GitHub — windmaple/autofinetune](https://github.com/windmaple/autofinetune) - [GitHub — google/tunix](https://github.com/google/tunix) - [GitHub — karpathy/autoresearch](https://github.com/karpathy/autoresearch) (cited reference for the paradigm) ## Frequently asked questions ### What is autofinetune at Google? It's a project presented on September 11, 2026 on the Google Developers Blog. It automates LLM post-training loops (supervised fine-tuning and GRPO reinforcement) using an agent that edits a script, launches jobs on Cloud TPU, measures a metric and commits or reverts according to the rules in `program.md`. ### What's the link with Tunix and Gemma? Tunix is Google's JAX-native post-training library used to run SFT and GRPO. The case studies cover FunctionGemma 270M (SFT) and Gemma 3 1B (mathematical GRPO on GSM8K). ### Do you need Gemini Flash 3.7 to use it? The post describes an orchestration with Antigravity CLI and Gemini Flash 3.7 as the agent of the outer loop. The training core remains Tunix on TPU. A team can draw inspiration from the pattern with another agent, but the official demonstration is set on this stack. ### What results does Google show? On SFT FunctionGemma, about twenty automated runs over a few hours with hill-climbing of function-call accuracy. On GRPO Gemma 3 1B, about forty runs over 2–3 days with roughly a 10% improvement in the published aggregated reward. These are example trajectories, not an independent public leaderboard. ### Why is this not a new frontier model? No new frontier weights are announced. Google is publishing a method and code to automate the exploration of hyperparameters and post-training prompts on already known TPU accelerators. ### Who is affected first? Teams already doing Gemma / Tunix post-training on Cloud TPU, who spend too much time on manual SFT or RL sweeps. Labs outside the Google stack will need to port the harness.