MiMo RL Environment Explorer

MiMo RL Environment Explorer

Browses 7,780 reinforcement learning environments from XiaomiMiMo, then runs a model on a task and shows how it is graded. An unofficial explorer for people who test or compare AI agents.

Made by FineEnvs on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: This Space is an unofficial explorer for the XiaomiMiMo MiMo-V2.6 RL dataset, which holds 7,780 reinforcement learning environments in five domains. You browse the tasks, open one to see what is inside, then run a rollout with a model you pick and watch it get graded. Hugging Face and the FineEnvs team built it.

Who it suits: Researchers and engineers who test AI agents on real tasks and want to see the grading as well as the answer. It also suits anyone who wants to read real agent traces.

Industry | Who uses it | How they use it | Example
Research | RL researchers | Browse tasks and read what each one contains | Open a Code task and look inside it
AI evaluation | Model evaluators | Run one task on several models and compare | Tick two to four rollouts and press Compare
Agent development | Agent builders | Try a model inside an agent framework on real tasks | Start a Webdev rollout and read its grade
Open data | Community readers | Learn from public rollouts | Browse the Community page

If you only want to chat with a model, you can skip it.

Where to use it: The explorer works in these places:

• A web browser: open the Space page on Hugging Face, browse tasks and run rollouts.
• Windows, macOS or Linux with Docker: build and run a copy on your own computer.
• Your Hugging Face account, which pays for the sandbox and the model calls.

Use case | Who | What they get back
Browse environments | Researchers | A task page showing everything inside the task
Run a rollout | Evaluators | A graded run of a model you chose, kept under Rollouts
Compare rollouts | Model teams | Settings, time, tokens, cost, grade checks and traces side by side

How to set it up: Browsing needs nothing. To run a rollout you need a few things:

• A Hugging Face account, signed in or with an access token pasted in
• A write token, or a fine-grained token with the Inference Providers and Jobs permissions
• Docker and Python with pip, only for a local copy

The steps download the Space, build a Docker image from it and run that image with port 7860 open.

1. Install the Hugging Face package.
2. Download the Space into a folder called MiMo-RL-Envs-Explorer.
3. Move into that folder.
4. Build the Docker image.
5. Run the image with port 7860 open.

pip install -U huggingface_hub
hf download FineEnvs/MiMo-RL-Envs-Explorer --repo-type space --local-dir MiMo-RL-Envs-Explorer
cd MiMo-RL-Envs-Explorer
docker build -t mimo-rl-envs-explorer .
docker run -it -p 7860:7860 mimo-rl-envs-explorer

The README also gives a shorter recipe for the source project's folder, which starts the app with uv and uses your Hugging Face token.

uv run uvicorn app.main:app

How to use it: Browse the list, open a task, then start a rollout with any model you like. Tasks that a model grades ask you to pick the judge. A rollout carries on after you close the page, and its entry waits under Rollouts.

Tool | What it does
Task page | Shows everything inside one task
Rollouts | Lists the runs you started
My rollouts | Shows only your own runs
Community | Shows public rollouts without who ran them
Compare | Puts two to four rollouts of one task side by side

The dataset has five domains:

• Code: 2,698 environments where the agent repairs a bug in a real repository, and tests it cannot see grade the work.
• Webdev: 2,093 environments where it builds a website from a written brief, and a vision model looks at the finished page.
• Cyber: 1,000 environments where it triggers a known crash again, and a locked server checks where the crash lands.
• Music: 1,000 environments where it writes a tune in ABC notation, scored on 18 human-likeness measures with no model.
• General: 989 environments of office-style knowledge work, graded by rule checks and rubric checks judged by a model.

Example: You open a Code task, pick a model and press run. The agent works in a sandbox built from the task's own Docker image, and hidden tests then grade whether the real issue was fixed.

For newcomers and experts: If you are new to this, open the Space page and browse the tasks first. You can read a whole task without signing in or spending anything. Run a rollout later, once you know which task you want to try.

If you know the field, a rollout starts an HF Sandbox and runs OpenCode in it. OpenCode calls the model you picked through Hugging Face Inference Providers. Then the task's own grader scores the run, using Xiaomi's reference code unchanged. While the agent works, the sandbox holds none of the grading material, and the agent cannot browse the web. Your token lives in an encrypted cookie, and the sandbox reaches models through a per-rollout proxy.

You pick a task and sign in
A sandbox starts from the task's Docker image
OpenCode runs the model you chose
The task's own grader scores the run
The rollout is saved, public or private

Good to know: The scores you see will differ somewhat from the ones Xiaomi published. The agent framework is OpenCode, the judge is the model you pick, and Webdev uses an evaluation-mode grader.

• License: apache-2, as the Space lists it.
• Cost to run: a Hugging Face account. A sandbox on CPU costs about $0.01 an hour, and model calls go through Inference Providers on your account.
• Data it reaches: your token, kept only in an encrypted cookie. Rollouts are public by default, shown without who ran them, and you can make any rollout private. Traces sit in a private bucket.
• Not for: matching Xiaomi's own training numbers, or tasks that need the web, since the agent has no web tools.

On Hugging Face, 13 November 2026

• Likes: 91
• Built with: docker
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/FineEnvs/MiMo-RL-Envs-Explorer
• Files: https://huggingface.co/spaces/FineEnvs/MiMo-RL-Envs-Explorer/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.