XiaomiMiMo Releases Agentic RL Training Environment and Dataset

September 27, 2026September 28, 2026

At a Glance

Values determined by this site’s code at collection time. Dates are JST.

Overview

XiaomiMiMo has released the dataset “XiaomiMiMo/MiMo-V2.6-RL-oss" for building reinforcement learning training environments for LLM agents (Agentic RL), along with a fork repository of the reinforcement learning framework verl named “XiaomiMiMo/verl". The target model specified is “MiMo-V2.6-Distill-Qwen-9B", and reproduction code for reinforcement learning environments across five domains (Code, Cyber, General, Visual, and Music) based on verl (0.9.0.dev) is presented.

This material outlines the specific environmental setup and implementation design of verifiers to reproduce the detailed training recipe described in Section 7 of the technical report “MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement" regarding scaling reinforcement learning methods for self-improvement.

Claims and Evidence

According to the presenter, the reproduction infrastructure for agentic reinforcement learning is structured as five types of RL environments combining domain-specific verification mechanisms ( verifiers) and task families. The correspondence of each environment provided within the fork repository of verl (0.9.0.dev) is as follows:

The characteristics of the task families and verification mechanisms in each domain are organized as follows:

• Code (Software engineering): Determines generation results using executable tests. The launch script is scripts/code/train.sh, and the configuration file corresponds to scripts/code/env.example.
• Cyber (Vulnerability reproduction): Performs judgment via rule-based checks. The launch script is scripts/arvo/arvo.sh, and the configuration file is assigned as scripts/arvo/arvo.env.example.
• General (Knowledge work): Evaluates using rubric-based judging. The launch script is scripts/general/general.sh, and the configuration file uses scripts/general/general.env.example.
• Visual (Web development): Verifies through visual grading. The launch script is scripts/design/webdev.sh, and the configuration file is scripts/design/webdev.env.example.
• Music (Symbolic music composition): Evaluates via rule-based checks. The launch script is scripts/design/music.sh, and the configuration file is scripts/design/music.env.example.

Additionally, two submodules placed under third_party/ in the repository are used as the agent execution infrastructure and trajectory collection mechanism to run these environments:

1. mimoagent (third_party/mimoagent-osr): A fork of mini-swe-agent, providing the agent harness, toolsets, execution environment, and grader. It is used in four domains: Code, Cyber, General, and Visual. Among these, in the three domains of Cyber, General, and Visual, mimoagent is driven by directly calling it from verl’s AgentLoop (recipes/{arvo,general,design}/agent_loop.py).
2. uni-agent (third_party/uni_agent): A fork of verl-project/uni-agent, responsible for the model gateway function and trajectory capture via TransferQueue. This submodule is used exclusively in the Code domain, replacing verl’s standard agent-loop manager to execute the mimoagent harness within the uni-agent session (recipes/code/mimoagent_runner.py).

According to the presenter’s materials, these two submodules do not directly import each other, and integration glue code is placed in the recipes/ directory. Note that the Music domain has an independent configuration that uses neither mimoagent nor uni-agent.

Prerequisites

To the extent explicitly stated in the materials, the conditions for reproducing the reinforcement learning pipeline presented are as follows:

• Target Model:
• XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B (released on Hugging Face)
• Reinforcement Learning Framework & Base Software:
• verl (version 0.9.0.dev)
• Submodules and External Components:
• third_party/mimoagent-osr (https://github.com/XiaomiMiMo/mimoagent, fork of SWE-agent/mini-swe-agent) – third_party/uni_agent (https://github.com/XiaomiMiMo/uni-agent, fork of verl-project/uni-agent)
• Training Dataset:
• XiaomiMiMo/MiMo-V2.6-RL-oss (released on Hugging Face under Apache-2.0 license) – Data format: Parquet format – Modality: Text, documents, images – Scale category: 1K to 10K entries – Configuration categories: 5 types (code, cyber, general, webdev, music) – Supported libraries: datasets, pandas, polars, mlcroissant
• Execution Container Environment:
• Docker image xiaomimimo/mimo-v2.6-rl-oss provided on Docker Hub
• Configuration Files:
• Environment variable template files corresponding to each domain (scripts/code/env.example, scripts/arvo/arvo.env.example, scripts/general/general.env.example, scripts/design/webdev.env.example, scripts/design/music.env.example)

What Can Be Reproduced Locally

The procedures and resources that can be executed and reproduced in a local environment using the public repository and dataset are published as follows:

Submodule Initialization

After obtaining the training code repository, initialize the submodules under third_party/ using the following command to make the agent environment and model gateway functional:

git submodule update --init third_party/mimoagent-osr third_party/uni_agent

Acquiring the Container Environment

As an execution environment, a pre-built Docker image is publicly available on Docker Hub.

docker pull xiaomimimo/mimo-v2.6-rl-oss

Launching Training for Each Domain

Before executing the launch script for each domain, it is necessary to refer to *.env.example in the same directory as each script and set the specified environment variables. Once configuration is complete, run the shell script corresponding to the target task.

• Code (Software engineering):
  After setting the variables in scripts/code/env.example, run scripts/code/train.sh. Internally, mimoagent is executed within a uni-agent session via recipes/code/mimoagent_runner.py, and trajectory acquisition is performed by TransferQueue.
• Cyber (Vulnerability reproduction):
  After setting the variables in scripts/arvo/arvo.env.example, run scripts/arvo/arvo.sh. Mimoagent is driven directly from verl’s recipes/arvo/agent_loop.py.
• General (Knowledge work):
  After setting the variables in scripts/general/general.env.example, run scripts/general/general.sh. Mimoagent is called directly from recipes/general/agent_loop.py.
• Visual (Web development):
  After setting the variables in scripts/design/webdev.env.example, run scripts/design/webdev.sh. Mimoagent is called directly from recipes/design/agent_loop.py.
• Music (Symbolic music composition):
  After setting the variables in scripts/design/music.env.example, run scripts/design/music.sh. Verification is performed via rule checks without using mimoagent and uni-agent submodules.

Detailed training recipes are noted to be described in Section 7 of the technical report “MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement“.

What Is Not Covered in the Material

This material and the README do not contain the following information:

• Hardware Requirements: Specific computer specifications such as GPU models required for training, recommended number of GPUs, VRAM capacity, and system memory capacity are not explicitly stated.
• Hyperparameter Recommendations: Specific setting values such as learning rate, batch size, parameters specific to RL algorithms like PPO, context length, and number of epochs are not directly described in the materials and require referencing environment variable configuration files (*.env.example) or external technical reports.
• Training Duration and Computational Cost: The calculation time required to complete reinforcement learning for each domain and the required time per node configuration are not documented.
• Benchmark Scores or Quantitative Verification Results: Comparison figures with baseline models or tables and measurement data of quantitative performance scores before and after reinforcement learning are not directly included in this material.

Related Articles

• Xiaomi Releases MiMo-V2.6-Pro-RL: 1.02T MoE Flagship
• Xiaomi Releases MiMo-V2.6-Flash-MOPD and Pro-MOPD Models
• SmolDataEnvs: RL Tasks for Small Model Optimization
• MiMo-V2.6-Distill-Qwen-9B-GGUF Vision-Language Model: 12GB+ VRAM

Sources

• XiaomiMiMo/MiMo-V2.6-RL-oss (Hugging Face Dataset)
• XiaomiMiMo/verl (GitHub Repository)
• XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B (Hugging Face Model)
• xiaomimimo/mimo-v2.6-rl-oss (Docker Hub)
• MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement (Technical Report)