The most valuable thing Xiaomi open-sourced in late September may not be the model itself, but the infrastructure that trained it. On September 22, 2026, Xiaomi released and open-sourced the MiMo-V2.6 omni-modal model family, consisting of two natively omni-modal models, Pro and Flash. But on September 21, the company created three repositories in a single day, laying out the reinforcement learning infrastructure behind that training run. The heaviest of the three is XiaomiMiMo/verl. Per a GitHub API snapshot taken September 27, 2026, the repository holds 465 stars and 49 forks, is written in Python, ships under the Apache-2.0 license, was created on September 21, and received its latest push on September 26.
First, the nature of the project, so the story does not drift: this is not an original Xiaomi RL framework. It is a fork. The upstream, verl-project/verl, had 23,650 stars in the same snapshot. It was initiated by ByteDance's Seed team and is maintained by the verl community; the open-source incarnation corresponds to HybridFlow, described in a paper accepted at EuroSys 2025, and the project migrated to the independent verl-project organization in January 2026. GitHub's fork relationship makes this unambiguous, and Xiaomi's own README says only this: the fork adds reproduction code for five RL environments on top of verl 0.9.0.dev, positioned as "Agentic RL training code for MiMo," with the training recipe in Section 7 of the technical report.
Standing on a giant's shoulders and adding the layer you know best is standard practice for a mature engineering team. What deserves a closer look is that layer: five RL environments covering five domains — Code for software engineering, Cyber for vulnerability reproduction, General for knowledge work, Visual for web development, and Music for symbolic music composition — each with its own verifier, launch script, and example environment file. These environments map directly onto the battlegrounds where MiMo-V2.6 was trained. According to Xiaomi's official figures as relayed by third-party coverage, the core of the recipe was large-scale Agentic RL with verifiable complex tasks as the reward signal: roughly six days of live RL, around 750,000 trajectories, a training context of one million tokens, and large batches on a fully asynchronous architecture. The environments determine what a model can practice; the verifiers determine what it actually learns.
Five Environments, One by One: The Verifier Is the Soul
Code, software engineering, verified by executable tests. Launch script: scripts/code/train.sh, with configuration in scripts/code/env.example. This is the most classic Agentic RL task family: the model modifies real code in a real repository, and the reward is decided by whether the test suite passes — no human judge, no second model acting as referee. Executable tests are the hardest currency among verifiers: the reward signal is crisp and there is little room to game it. Third-party coverage notes that MiMo-V2.6 showed clear out-of-sample gains on DeepSWE v1.1, and this is the environment family behind that.
Cyber, vulnerability reproduction, verified by rule checks. Launch script: scripts/arvo/arvo.sh. The model reproduces known vulnerabilities in a controlled environment, and the reward is determined by rules: whether the trigger condition holds and whether the reproduction path works. Turning vulnerability reproduction into an RL environment is rare in the open-source framework world, and it explains why security appears as a distinct line in Xiaomi's capability list.
General, knowledge work, verified by rubric-based judging. Launch script: scripts/general/general.sh. These tasks have no single correct answer — writing a research brief, producing an analysis — so the verifier shifts from right-versus-wrong to scoring against itemized rubric criteria. Rubric judging is the most design-intensive of the five: the finer the criteria, the steadier the reward signal.
Visual, web development, verified by visual grading. Launch script: scripts/design/webdev.sh. The output of a web development task has to be judged by how it renders: whether the layout is right, how faithfully the styles match, whether the page is usable. Moving the verifier from text space into visual space means the model's "eyes" are wired into the reward signal — the most easily neglected and hardest-to-get-right part of omni-modal training, because the standard for visual quality is far fuzzier than a unit test.
Music, symbolic music composition, verified by rule checks. Launch script: scripts/design/music.sh. Music theory rules are naturally suited to formal checking: intervals, harmony, rhythm, and structure can all be judged mechanically. Folding composition into the verifiable-reward world is the most imaginative entry in this matrix — the implied claim is that any domain where you can write down rules is fair game for Agentic RL.
Read together, the design philosophy behind the five verifiers is clear: formalize what can be formalized (executable tests, rule checks) and structure what cannot (rubric scoring, visual grading). The coverage runs from plain-text code, through vision, into symbolic music — filling out the capability spectrum of an omni-modal model almost exactly.
Why dwell on verifiers? Because Agentic RL succeeds or fails on them. RL is driven by reward, and models will raise the reward rather than do the task well — that is reward hacking. Xiaomi's official account lists suppressing it as a pillar of stability design: reward engineering, adversarial evaluation, anomaly detection, and cross-checked verifiers. This matrix effectively open-sources the first line of defense — where the reward comes from. Understanding the verifier design is worth far more than understanding the launch scripts.
What Else Is in the Repo: Submodules, Images, and the Training Model
The environment table alone is not enough to run anything. The README describes three pieces of supporting infrastructure.
First, two submodules under third_party, initialized with git submodule update. The first is mimoagent (third_party/mimoagent-osr, itself a fork of mini-swe-agent from the SWE-agent team), which provides the agent harnesses, tools, execution environments, and graders. Code, Cyber, General, and Visual all run on it; Cyber, General, and Visual drive it directly from a verl AgentLoop, with the glue code in the recipes directory. The second is uni-agent (third_party/uni_agent, a fork of verl-project/uni-agent), which provides a model gateway and TransferQueue trajectory capture. Only Code uses it: it replaces verl's agent-loop manager and runs the mimoagent harness inside a uni-agent session. The README is explicit that the two submodules do not import each other and that Music uses neither — a sign of carefully drawn dependency boundaries.
Second, a Docker image hosted on Docker Hub as xiaomimimo/mimo-v2.6-rl-oss. RL environments run test suites, browsers, and audio processing, with sprawling dependency chains; the official image flattens that setup for you, which beats a ten-page installation guide by a wide margin.
Third, the training model, MiMo-V2.6-Distill-Qwen-9B, hosted on Hugging Face. The starting point of the full reproduction path is not the flagship model, whose parameter count has not been disclosed, but a 9-billion-parameter distillation. For teams with limited compute, that is good news: five environments plus a 9B model pushes the reproduction threshold down to a realistic scale.
The usage barrier is a single convention: each launch script reads the variables listed in the env.example file next to it, so set those before launching. It keeps the configuration explicit — no hidden settings buried in the code, and no guessing about where to tune.
Dataset and Technical Report: The Reproduction Path Is Complete
Plenty of training frameworks are open-sourced with code alone. The real weight of this repository is that Xiaomi released all three things a reproduction requires: environment code, training data, and the training recipe.
The training dataset, MiMo-V2.6-RL-oss, is openly downloadable on Hugging Face. Most teams open-source a framework with the data missing, and without data the environments are an empty shell — you never see what the tasks looked like or how the rewards were aligned. Open data also lets anyone run controlled experiments against this exact set to test which parts of the design actually matter.
The technical report PDF sits in the XiaomiMiMo/MiMo-V2.6-Pro-RL repository on Hugging Face, under the filename MiMo_V2_6_technical_report.pdf. The training recipe is in Section 7, and the report is titled "MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement." How the environments are configured, how rewards are computed, and how reward hacking is suppressed — the report carries the written version, and the repository carries the executable version.
Assemble the three pieces and a complete reproduction path appears: clone the repository, initialize the submodules, pull the Docker image, download the dataset, set parameters per Section 7 of the report, and run the launch script for the environment you care about. For teams doing Agentic RL, that is worth far more than one more benchmark model. Peers open-sourced results; Xiaomi open-sourced the process.
A Same-Day Bundle: Three Repositories, One System
The verl fork is not the only repository created on September 21. Two companions shipped the same day: uni-agent, at 14 stars and Apache-2.0, a long-horizon agent training framework; and mimoagent, at 30 stars and MIT licensed, self-described as a minimal agent that solves GitHub issues in 100 lines. Three repositories, one creation date, clear roles: the verl fork is the training hub, uni-agent handles long-horizon training and trajectory capture, and mimoagent acts as the lightweight executor. This is a complete training infrastructure bundle, not scattered open-sourcing.
One detail worth noticing is uni-agent's fork identity: it is a fork of verl-project/uni-agent, a unified agent framework built on top of verl and released in May 2026. Xiaomi reused a component from the upstream ecosystem inside its own environments, then gave the modified version back as an independent repository — a small closed loop.
Cold Thoughts: Do Not Rush to Declare It Beats the Upstream
Four things warrant a level head.
First, the ecosystem is very young. Against the upstream's 23,650 stars, 465 makes this fork a small project, with an issue pool, community discussion, and third-party tutorials all hovering near zero. Anyone adopting it as a production tool will be breaking their own trail, though early entry also means fast feedback and a real chance to influence the project's direction.
Second, the divergence details are unknown. GitHub compare shows the fork at ahead_by 0 and behind_by 16 relative to upstream, and the upstream was still committing actively around September 26 — several batches of upstream commits are not yet merged into the fork. What Xiaomi changed inside the fork, and whether they intend to contribute it back, is not stated in the README and cannot be verified. Claims that the Xiaomi version is ahead of upstream by some number of commits are neither provable nor publishable; the only verifiable facts are the five environments' reproduction code in the README.
Third, the upstream has not slowed down. RL-Insight released in July 2026, verl-SpeCo pre-released in June, uni-agent and verl-Omni in May — the cadence on the main verl repository shows no sign of easing, and whether the fork keeps up with upstream evolution is a question the Xiaomi team will have to answer.
Fourth, the licensing details. The GitHub API metadata lists Apache-2.0; one caveat applies: treat the LICENSE file in each repository as the final authority.
The Bottom Line
Return to the opening question: why open-source the infrastructure behind a model? For Xiaomi, it is ecosystem positioning around MiMo-V2.6 — letting teams doing Agentic RL step directly onto ground Xiaomi has already scouted. For the industry, it fills in a rarely seen piece of the puzzle: the RL environment design and verifier schemes behind a frontier model, laid out for inspection. Five environments, verifier designs, a dataset, a technical report, and official Docker images — the chain is complete. The shortcomings are equally clear: an early ecosystem, undisclosed divergence details, and an upstream still running fast. Teams considering adoption should read Section 7 of the technical report first, then decide whether a private fork is worth the effort.
Two points of clarification for readers navigating related coverage: Xiaomi's June release, MiMo-Code, is a terminal coding assistant — a product — and entirely separate from the training infrastructure covered here. For more angles on the same story, this batch also includes a MiMo-V2.6 hotspot piece, an Agentic RL frameworks comparison, and a MiMo-V2.6 integration SOP. One editorial note on division of labor: the site's general LLM fine-tuning SOP covers the SFT/LoRA route, while this article focuses on Agentic RL training infrastructure; the two are complementary. For a comparison of other open flagships, the Kimi K3 open-weights piece is a useful companion read.
FAQ
Q1: Is XiaomiMiMo/verl an original Xiaomi RL framework? A1: No. It is a fork of verl-project/verl (HybridFlow, 23,650 stars, initiated by ByteDance's Seed team and community-maintained). Xiaomi added reproduction code for five Agentic RL environments on its verl 0.9.0.dev base, and the README's own positioning reads "Agentic RL training code for MiMo."
Q2: What verifier does each of the five environments use? A2: Code uses executable tests, Cyber uses rule checks, General uses rubric-based judging, Visual uses visual grading, and Music uses rule checks. The launch scripts are scripts/code/train.sh, scripts/arvo/arvo.sh, scripts/general/general.sh, scripts/design/webdev.sh, and scripts/design/music.sh.
Q3: What materials are needed to reproduce, and where do they live? A3: Four items: the repository itself (including the mimoagent and uni-agent submodules under third_party), the Docker image xiaomimimo/mimo-v2.6-rl-oss, the MiMo-V2.6-RL-oss training dataset on Hugging Face, and the technical report PDF with the recipe in Section 7. The training model is MiMo-V2.6-Distill-Qwen-9B.
Q4: What is the license, and what should you check before commercial use? A4: The GitHub API metadata lists Apache-2.0; uni-agent is also Apache-2.0 and mimoagent is MIT. Before commercial use, verify each repository's LICENSE file itself as the final authority.
Q5: Is it better than upstream verl, and which should you use? A5: There is no basis for comparison. GitHub compare shows ahead_by 0 and behind_by 16, the divergence details are unverified, and upstream is still actively iterating. For a general RL post-training framework, use upstream; to reproduce the five Agentic RL environments, use this fork.