Xiaomi Releases MiMo-V2.5 Open-Source Model Series

aiai-modeling 4 posts · 4 accounts

Xiaomi released two open-source AI models, MiMo-V2.5-Pro and MiMo-V2.5, and said the new series is aimed at agentic tasks, software engineering and multimodal workloads. The company described MiMo-V2.5-Pro as its strongest model yet, citing benchmark scores of 57.2 on SWE-bench Pro, 63.8 on Claw-Eval and 72.9 on τ3-Bench, while saying MiMo-V2.5 offers native omnimodal capabilities, image and video understanding, a 1M-token context window and pro-level agent performance at roughly half the cost.

The models are available in Hermes Agent through Nous Portal and OpenRouter. Xiaomi also said the series was tuned for token efficiency, with MiMo-V2.5-Pro using 42% fewer tokens than Kimi K2.6 at the same ClawEval score and MiMo-V2.5 using nearly half the tokens of Muse Spark.

From the sources (4 posts)

@teortaxestex

Half a year ago, AGI Agents were proudly booking flights and hotel rooms. Now even second-tier labs are flexing with "implement a SysY compiler in Rust" or graduate-level analog EDA tasks instead. This will be open sourced btw do you feel

@teknium

Xiaomi's MiMo-V2.5 and MiMo-V2.5-Pro are both now available in Hermes Agent through Nous Portal and OpenRouter! Just `hermes update`!

@nousresearch

RT @Teknium: Xiaomi's MiMo-V2.5 and MiMo-V2.5-Pro are both now available in Hermes Agent through Nous Portal and OpenRouter! Just `hermes…

@testingcatalog

Xiaomi released new MiMo-V2.5-Pro and MiMo-V2.5 open-source models. > MiMo-V2.5-Pro, the strongest model yet. > MiMo-V2.5, native omnimodal with strong agentic capabilities.

Market impact

42 medium

NVDA MSFT GOOGL SOXX QQQ

Why it matters: Another capable open-weight, cost-efficient frontier-class model (agentic/SWE benchmarks, ~half the cost, 42% fewer tokens than Kimi K2.6) incrementally pressures the closed-lab margin thesis and the AI-capex return case, but as a second-tier lab in an already-crowded open-weight field it is a diluted DeepSeek-style read-through rather than a fresh shock.

At publish time: Headline/summary alone flag it: open-source frontier-class release with concrete SWE-bench/agentic benchmark scores plus explicit token-efficiency and ~half-cost claims vs Kimi K2.6.

model confidence: low

What happened next

1–2 weeks later

fizzled moved up

Bearish semis read-through failed; NVDA rallied from ~$177 to $217 by Apr 27 and $235 by mid-May. Follow-ups are routine open-weight releases, no shock.

based on follow-up coverage

Follow-up coverage

No follow-up coverage matched this story.

Cloud Providers Reserve Nvidia GPUs, Raising Waits for AI Startups Saudi Arabia's Arab Light Trades at Record Premium to Brent