MiMo-V2.6-Pro

weights

open(MiMo-V2.6-Pro-RL and MiMo-V2.6-Pro-MOPD on Hugging Face, ungated safetensors)

data

closed(no pretraining corpus released or named)

code

partial(inference support and RL post-training code (a verl fork with five environment recipes)

license

MIT(HF card on both V2.6 Pro checkpoints)

rl-environments

released(MiMo-V2.6-RL-oss on Hugging Face under Apache-2.0: code, cybersecurity, knowledge-work, web-development and music environments, graded by tests, rule checks, rubrics and visual grading)

agent-harness

released(mimoagent (MIT, a fork of mini-swe-agent) supplies the harnesses, tools and graders, and uni-agent (Apache-2.0, a fork of verl-project/uni-agent) the model gateway and trajectory capture, both as submodules of the verl fork)

Xiaomi publishes the MiMo-V2.6-Pro weights under MIT, along with reinforcement-learning code and environments that reproduce its training recipe on a distilled 9B model. The pretraining corpus is not released and no pretraining code ships, so the model is open weights rather than fully open.

• https://api.github.com/repos/XiaomiMiMo/mimoagent recorded 2026-09-27

  `fork: true` with parent `SWE-agent/mini-swe-agent`, license `MIT`

• https://api.github.com/repos/XiaomiMiMo/uni-agent recorded 2026-09-27

  `fork: true` with parent `verl-project/uni-agent`, license `Apache-2.0`

• https://api.github.com/repos/XiaomiMiMo/verl recorded 2026-09-27

  `fork: true` with parent `verl-project/verl`, license `Apache-2.0`, created 2026-09-21

• https://huggingface.co/api/models/XiaomiMiMo/MiMo-V2.6-Pro-MOPD recorded 2026-09-27

  card license `mit`, `gated: false`, `private: false`, created 2026-09-27, with the same weight-shard layout as Pro-RL

• https://huggingface.co/api/models/XiaomiMiMo/MiMo-V2.6-Pro-RL recorded 2026-09-27

  card license `mit`, `gated: false`, `private: false`, created 2026-09-21; safetensors weight shards and MiMo_V2_6_technical_report.pdf in the file list

• https://huggingface.co/datasets/XiaomiMiMo/MiMo-V2.6-RL-oss/raw/main/README.md recorded 2026-09-27

  'license: apache-2.0', 'RL training environments for LLM agents' across Code, Cyber, General, Visual and Music, with executable tests, rule checks, rubric-based judging and visual grading

• https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL recorded 2026-09-27

  card licensed MIT, '1.02T total / 42B activated parameters', a Downloads table listing the Pro and Flash checkpoints; no training corpus named

• https://raw.githubusercontent.com/XiaomiMiMo/verl/mimo-oss/README.md recorded 2026-09-27

  'Agentic RL training code for MiMo' and 'This fork adds reproduction code for five RL environments'; its submodules are mimoagent ('agent harnesses, tools, execution environments and graders') and uni-agent ('model gateway and TransferQueue trajectory capture'), and the training model is MiMo-V2.6-Distill-Qwen-9B

Adoption is read from Hugging Face downloads of MiMo-V2.5-Pro, the only checkpoint declared when it was measured; the V2.6 checkpoints are newer than that reading.

• https://huggingface.co/api/models/XiaomiMiMo/MiMo-V2.5-Pro recorded 2026-08-12

  59,525 downloads in the trailing 30 days for XiaomiMiMo/MiMo-V2.5-Pro

Artificial Analysis ranks MiMo-V2.6-Pro first of 115 models on its Intelligence Index, two places above Kimi K3, and calls it among the leading open-weights models of its size in intelligence. Xiaomi's own agentic benchmarks put it close to Claude Opus 5 on software-engineering and computer-use tasks, though Xiaomi ran those itself.

• https://artificialanalysis.ai/models/kimi-k3 recorded 2026-09-27

  'Kimi K3 (max) scores 44 on the Artificial Analysis Intelligence Index', ranked #3 of 115 in the model summary

• https://artificialanalysis.ai/models/mimo-v2-6-pro recorded 2026-09-27

  'MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index' (v4.3.2), ranked #1 of 115 in the model summary, and 'amongst the leading models in intelligence' against 'other open weight models of similar size'; released September 21, 2026

• https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL recorded 2026-09-27

  Xiaomi's evaluation table: DeepSWE v1.1 71.9, Terminal Bench 2.1 89.9 and OSWorld-Verified 82.0 for MiMo-V2.6 Pro, against 74.0, 89.1 and 83.4 for Claude Opus 5