Xiaomi: every model, side by side
Builds the open-weight MiMo models for long-context, reasoning, coding, and agent tasks.
- MiMo models emphasize long-context reasoning and agent scenarios.
- Mixture-of-Experts, tuned for performance per parameter.
- Total models
- 7
- Text models
- 7
- Active period
- Dec 2025 to Sep 2026
Xiaomi models and answers
MiMo-V2.6-Flash
Sep 2026Xiaomi MiMo-V2.6 Flash accepts text, image, audio, and video inputs and produces text. Its OpenRouter endpoint supports a 1M context window with lower token pricing than the Pro variant.
conversation · reasoning · tool-use
MiMo-V2.6-Pro
Sep 2026Xiaomi MiMo-V2.6 Pro accepts text, image, audio, and video inputs and produces text, with a 1M context window on OpenRouter.
conversation · reasoning · tool-use
MiMo-V2.5
Apr 2026Xiaomi's native omnimodal MiMo-V2.5, roughly half the inference cost of the Pro tier and ahead of MiMo-V2-Omni on image and video understanding. Its 1M context takes whole documents, long conversations and complex task state in one pass, which is what makes it usable inside agent frameworks.
conversation · reasoning · code-generation · analysis · agentic-tool-use
MiMo-V2.5-Pro
Apr 2026Xiaomi's largest MiMo-V2.5, ranked near the top on ClawEval, GDPVal and SWE-bench Pro. It runs professional tasks that would take a human expert days, chaining more than a thousand tool calls without supervision, across a context of up to 1M tokens.
conversation · reasoning · code-generation · analysis · agentic-tool-use · planning
MiMo-V2-Pro
Mar 2026Xiaomi's largest MiMo model, over 1T total parameters with a 1M context, tuned for agent frameworks like OpenClaw. Xiaomi places it in the top tier on PinchBench and ClawBench, near Opus 4.6. The intended job is orchestrating workflows and production engineering tasks rather than answering single questions.
conversation · reasoning · code-generation · analysis · agentic-tool-use · planning
MiMo-V2-Omni
Mar 2026An omni-modal model that takes image, video and audio in one architecture. Adds visual grounding, multi-step planning, tool use and code execution for tasks that span modalities.
conversation · reasoning · code-generation · analysis · agentic-tool-use
MiMo-V2-Flash
Dec 2025Xiaomi's open-source MiMo-V2-Flash, a mixture-of-experts with 309B total and 15B active on a hybrid attention design. 256K context, thinking toggle. Xiaomi reports it top open-source on SWE-bench Verified and SWE-bench Multilingual, near Claude Sonnet 4.5 at about 3.5% of the cost.
conversation · reasoning · code-generation · analysis · agentic-tool-use