MiMo V2.5 Pro
Editorial notes
Beta publica 22 abril 2026. 1.02T params MoE / 42B activos. Hybrid attention 6:1 sliding-window:global, 128 sliding window. Reduce KV-cache ~7x. Sucesor directo de MiMo V2 Pro. Open-weight, weights en huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro. Pricing $1.00/$3.00 per MTok (256K). Scores del modelo FINAL/instruct (blog oficial mimo.xiaomi.com, en imagen post_training_evaluation): SWE-bench-Verified 78.9 (con scaffold propio OpenClaw), SWE-bench-Pro 57.2 (tercero llm-stats lo confirma 0.572), Terminal-Bench 2.0 68.4. Otros vendor fuera del core-set: ClawEval 64% Pass^3, GDPVal-AA Elo 1581, MiMo Coding Bench 73.7. Tercero: AA Intelligence Index 42 (#4 de 90), HLE 48 (the-decoder, sin confirmar con/sin tools). TRAMPA Base-vs-Final: GPQA-D 66.7 / MMLU-Pro 68.5 / AIME 24&25 37.3 / LiveCodeBench 39.6 son del modelo BASE (pre-training), NO del instruct — no cargados. No alcanza el gate: los 3 benchmarks vendor del core-set son todos coding/agentic; Xiaomi NO publico un reasoning (GPQA-D/HLE) del modelo final.
Spec sheet
- Company
- Xiaomi
- Country
- CN
- Type
- hybrid
- Release
- 2026-04
- Context
- 1.0M tokens
- Slug
- mimo-v2-5-pro
Benchmarks (3)
Cite this model
BibTeX · APA
BibTeX
@misc{frontier-mimo-v2-5-pro,
title = {MiMo V2.5 Pro},
author = {{Xiaomi}},
year = {2026},
note = {Frontier Benchmarks AI atlas. Accessed 2026-06-17},
url = {https://frontierbenchmarks.com/models/mimo-v2-5-pro}
} APA
Xiaomi (2026). MiMo V2.5 Pro [Large language model]. Frontier Benchmarks AI. Retrieved 2026-06-17, from https://frontierbenchmarks.com/models/mimo-v2-5-pro
Citation reflects the atlas page, not the original model paper. For the paper, see the "Resources" section above.