Model components / Base / pretrained models

MiMo-V2.5-Pro

Xiaomi's open-weights mixture of experts, released April 2026 with about 1.02 trillion total and 42 billion active parameters across 70 layers and eight experts per token. It interleaves sliding-window and global attention at a six-to-one ratio, uses three-layer multi-token prediction, and reaches a one-million-token context. It targets agentic, software-engineering and long-horizon tasks.

An older MiMo repository carries different terms from the model card, and the card is authoritative here. Xiaomi's product page no longer carries the evaluation tables and points to the card, so the figures are read there. Verified 2026-08-13 via the model card, the product page and the MiMo repository.

Openness

3 medium confidence

3.0

weights: open(MIT per HF card)
data: closed
code: open
license: MIT(HF card

Open weights under a permissive license - MIT on the Hugging Face card, though the older GitHub repository shows Apache-2.0 - with no open training data, so open weights rather than open source.

https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro recorded 2026-08-13
MIT, 1.02T/42B MoE, 1M context, benchmark tables; no training corpus named on the card
https://huggingface.co/api/models/XiaomiMiMo/MiMo-V2.5-Pro recorded 2026-08-13
card license `mit`, `gated: false`, `private: false`, 61,546 downloads and 736 likes.
https://api.github.com/repos/XiaomiMiMo/MiMo recorded 2026-08-13
`private: false`, `archived: false`, license `Apache-2.0`, 2,294 stars, description "MiMo: Unlocking the Reasoning Potential of Language Model - From Pretraining to Posttraining". This is the older repo whose Apache-2.0 the license note contrasts with the card MIT.

Adoption

2 medium confidence

2.0

59,525 Hugging Face downloads in the trailing 30 days for XiaomiMiMo/MiMo-V2.5-Pro, which bands at level 2 (10K-100K) on the model adoption scale.

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

Capability

4 medium confidence

4.0

A frontier open-weights model built for efficiency. The score rests on the benchmarks Xiaomi publishes: its own product page defers to the Hugging Face model card, where MMLU 89.4, MMLU-Pro 68.5, GSM8K 99.6 and HumanEval+ 75.6 all appear. The LMArena rank of roughly #29 is reported by mirrors with no citable source behind it, so it colors the picture without carrying the score.

https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro recorded 2026-08-13
MMLU 89.4, MMLU-Pro 68.5, GSM8K 99.6 and HumanEval+ 75.6 on the official model card
https://mimo.xiaomi.com/mimo-v2-5-pro recorded 2026-08-13
official architecture and Apr 27 2026 release announcement. The benchmark figures are no longer readable on this page - it now points to the Hugging Face model card for the evaluation tables.

Verified 2026-08-12