Xiaomi Open-Sources CodeMidas: Mining 5,545 RL Training Tasks from Existing Repository Code

Xiaomi's MiMo team released and open-sourced CodeMidas on September 22. The name inverts the King Midas legend: instead of turning things to gold, the gold in code needs to be mined out. The pipeline converts features that are already implemented in open-source GitHub repositories into executable reinforcement learning training environments. After training with these environments, MiMo-V2.5 improved by 11.7% on DeepSWE, 17% on ProgramBench, and 8.5% on Terminal-Bench v2.1.

Why This Approach Works

RL training for coding is bottlenecked by two questions: where do tasks come from, and who judges correctness?

Most prior work starts from issues or commits. The problem is that already-implemented functionality in repositories far outnumbers filed tickets. A parse_url function, a sorting algorithm, a set of config-parsing logic — these are behaviors written during development and rarely changed afterward. To a model, they are natural tasks: input to output, testable, runnable.

CodeMidas's core idea is closed-loop iteration:

1. Explore behavior — an agent explores implemented features in open-source repositories. 2. Write behavior specs — inputs, outputs, side effects. 3. Build tests from original execution — no hand-written prompts needed. 4. Filter reliable tasks via repeated rollout unwinding. 5. Train — MiMo-V2.5 with GRPO on the curated task set.

Each step has hard acceptance criteria: tests must actually run, random shortcut solutions must fail, and correct answers must be findable across multiple rollouts. The pipeline requires neither a pre-stored issue database nor human-written prompts.

The Scale of the Data

The final training set covers 3,185 open-source repositories, 23 programming languages, 15 technical domains, and 5,545 tasks in total.

The 17% jump on ProgramBench is the largest single improvement that benchmark has seen since its release. The 8.5% gain on Terminal-Bench v2.1 came in the terminal-operation category, a scenario where training data has historically been sparse.

Ablation results are straightforward: more high-quality tasks yield better performance, with no obvious saturation in the curve.

Why This Route Goes Further Than "Grinding Issues"

Training coding agents from issues hits two ceilings:

• Task coverage: implemented features in open-source repositories outnumber reported bugs by orders of magnitude.
• Verification reliability: tests derived from issue descriptions often require manual maintenance. CodeMidas generates tests directly from running the original code, and can filter out tasks that only pass by chance across multiple rollouts.

This paradigm treats repositories as a library of training material, not as projects to fix. The boundary of task sources expands to all code ever written, not just reported problems.

What Changed in Behavior

Trajectory analysis shows post-RL MiMo-V2.5 improved on two fronts:

1. More thorough exploration: longer traversal paths from repository entry points, digging into subdirectories for implementation details instead of stopping at READMEs and interface signatures. 2. More diverse self-verification: running tests, checking against expected outputs, comparing differences across candidate solutions — rather than relying on a single check.

This closes the loop with RL task design: the task definitions demand deeper trajectories, and the data teaches the behavior more thoroughly.

Unavoidable Limitations

• Implemented behavior ≠ complete specification: automated tests may miss edge cases such as exception branches or concurrency races.
• Repeated rollouts ≠ strict quality control: no amount of reruns replaces human sampling audits.
• Task distribution skews toward real-world code: few tutorial-style exercise repos survive filtering. The remaining tasks are mostly engineering implementations — friendlier to engineering ability, less helpful for competitive programming.

CodeMidas solves the supply-side problem of training tasks, not the problem of verification correctness. Pushing this route toward serious benchmarks will require adding human effort on the verification side.

The Simultaneously Open-Sourced "Training Livestream"

Xiaomi also released MiMo-V2.6 full-modal model weights (Pro / Flash / Distill-Qwen-9B), an end-to-end RL training framework, and 7,000+ high-quality RL task environments. The effort, led by Luo Fuli (former DeepSeek researcher who joined Xiaomi in November 2025 to lead MiMo), has been described as a "training livestream" and reached 500+ discussion points on Hacker News. Everything is MIT-licensed for commercial use and fine-tuning.

Model highlights:

• MiMo-V2.6-Pro: 1.02T total parameters / 42B activated, sparse MoE architecture, full-modal (text/image/video/audio), 1M token context.
• MiMo-V2.6-Flash: 309B total parameters / 15B activated.
• Distill-Qwen-9B: SWE-bench Verified 61.1 → 66.2; Terminal-Bench 2.1 37.1 → 52.8.

The training process was disclosed in detail: 30 RL steps, each with 1,568 prompts × 16 rollouts = 25,088 samples, consuming 3.5–3.7 billion tokens per step, completed in 6 days. On DeepSWE v1.1, Flash rose from 48.8 to 65.68 and Pro from 58.4 to 72.57.

> Sources: Hugging Face Daily Papers, Xiaomi MiMo public pages, NetEase/AGI HUNT coverage from 9-22, CCTest summary.

References

1. Xiaomi MiMo CodeMidas project page and open-source repository (9-22, Hugging Face Daily Papers): 3,185 repos / 5,545 tasks / 23 languages / 15 domains; +11.7% / +17% / +8.5%. 2. NetEase Tech, "Xiaomi MiMo-V2.6 Turns LLM Training into a Public Experiment" (9-22): 6 days, 30 RL steps, 25,088 samples per step, Pro/Flash evaluation curves, Luo Fuli background. 3. CCTest, "CodeMidas Builds Coding RL Environments from Source Code" (9-22): four-stage pipeline, behavior spec + test construction + multi-rollout filtering definitions. 4. AGI HUNT 9-22 Daily: roundup of CodeMidas and the concurrent Xiaomi MiMo-V2.6 open-source releases.