Xiaomi MiMo-V2.6 Leak Puts Open Models Near Frontier Rivals

Xiaomi's next open agent models are showing up in a leaked benchmark table, and the numbers say the gap to the closed frontier has mostly closed. The MiMo-V2.6 series comes in two versions, Pro and Flash, and both are compared against MiMo-V2.5, Claude Opus 5 and GPT-5.6 Sol.

The catch is right at the top: this is a leaked comparison table, not a Xiaomi announcement. Nothing here is confirmed, and there is no word yet on a release date, pricing or availability.

What Happened

• DeepSWE v1.1: MiMo-V2.6 Pro scores 71.9, up from 19.0 on MiMo-V2.5 Pro. That is a generational jump, not a tweak.
• AutomationBench v1.0.6: 53.1 on Pro and 52.3 on Flash, up from 16.0 last generation. Pro lands ahead of Claude Opus 5 (50.3) and GPT-5.6 Sol (45.8).
• JobBench: 62.0 on Pro. That beats GPT-5.6 Sol (45.4) and sits just behind Opus 5 (65.7).
• Terminal Bench 2.1: 89.9 on Pro, the top score in the table, ahead of Opus 5 (89.1) and GPT-5.6 Sol (88.8).
• CyberGym, which covers autonomous vulnerability handling: 95.1 on Flash and 94.0 on Pro, so the cheaper model wins this one.
• Multimodal visual coding: 72.3 on Pro and 71.5 on Flash. That clears Opus 5 (70.0) but stays under GPT-5.6 Sol (73.4).

Why It Matters

If these numbers hold up once the models actually ship, agent work like running terminal tasks, automating workflows and patching vulnerabilities stops being something only paid APIs do well. An open-weight model from a consumer hardware company would let you point your dev stack at your own weights instead of somebody else's endpoint, which changes what that work costs you.

The honest read is that a leaked table is a claim, not a result. Wait for the official release and independent runs before you rip out the API you already trust. Keep the numbers in mind, but test on your own workload first.