Xiaomi MiMo-V2.5-Pro releases with 57.2 SWE-Bench Pro, 1M context, and OpenRouter access
Xiaomi’s MiMo-V2.5-Pro and MiMo-V2.5 arrived with million-token context windows, stronger coding and agentic claims, and immediate access through OpenRouter plus agent harnesses. The rollout adds another low-cost Chinese frontier model that engineers can route into coding workflows without waiting for a proprietary IDE deal.
TL;DR
- Xiaomi’s OpenRouter launch brought MiMo-V2.5 and MiMo-V2.5-Pro online with a stated focus on “long running agent tasks,” stronger coding, and 1M-token context windows.
- The benchmark card in testingcatalog’s roundup positions MiMo-V2.5-Pro at 57.2 on SWE-Bench Pro and 68.4 on Terminal-Bench 2.0, roughly alongside Claude Opus 4.6 on several agent tasks.
- Tooling support landed immediately: opencode’s update added both models in Go with “price unchanged,” and Teknium’s post says Hermes Agent can access them through Nous Portal and OpenRouter.
- Early long-context testing is less clean than the launch spec: according to Context Arena, MiMo v2 Pro held up through 64k but the team “couldn’t reliably run the 1M bin” because of context-length enforcement and tokenizer issues.
What shipped and where it already plugs into agent workflows
Xiaomi’s release matters less as a model-card event than as a same-day routing event. OpenRouter says both models are live now, with MiMo-V2.5 pitched as omnimodal and MiMo-V2.5-Pro as the stronger coding model; the OpenRouter model page is linked from the launch post.
The launch image shared in testingcatalog’s roundup gives the practical shape of the release:
- MiMo-V2.5-Pro scores 57.2 on SWE-Bench Pro versus 56.1 for MiMo-V2.5.
- On Terminal-Bench 2.0, Pro posts 68.4, ahead of Claude Opus 4.6 at 65.4 and just behind GPT-5.4 at 75.1.
- On t3-bench, Pro reaches 72.9, matching GPT-5.4 in the chart.
- Humanity’s Last Exam stays lower at 48.0 without tools, which keeps the release centered on coding and agent work more than frontier reasoning.
Availability also arrived in existing agent harnesses instead of a Xiaomi-only surface:
- opencode’s update says both models are available in Go, with v2.5 labeled “multimodal” and Pro “built for coding.”
- Teknium’s post says Hermes Agent users can pull them in with
hermes updatevia Nous Portal and OpenRouter. - The long-context caveat from Context Arena is operational: the advertised 1M window exists, but external eval infrastructure still hit provider-side context errors before it could validate a full 1M-bin run.