← Back to feed

Xiaomi’s MiMo V2.6 Pro and MiMo V2.6 Flash joined Arena’s Agent Arena leaderboard on October 15, 2026. Both carry MIT licenses. Both are fully open-weight. And both are now competing directly against proprietary frontier systems from Anthropic, OpenAI, and Google on agentic task evaluation.

The addition matters not because MiMo V2.6 is a new model — Xiaomi released the V2.6 generation in September 2026, and the Pro variant has appeared on the Code Arena top 10 and tied Grok 4.7 at Artificial Analysis Intelligence Index 46. What’s new is the leaderboard: the Agent Arena specifically evaluates agentic performance, not just instruction-following quality or code generation.

Model Specs

MiMo V2.6 Pro is a 1.02 trillion-parameter MoE model with 42 billion active parameters. It runs under MIT license, meaning it can be deployed commercially, fine-tuned, and redistributed without restriction.

MiMo V2.6 Flash is the smaller configuration: 309 billion total parameters with 15 billion active. Also MIT licensed. On SWE-bench Verified (Thinking mode), Flash scores 78.6%. The Non-Thinking variant scores 73.7%.

For the Pro variant, available benchmark data includes DeepSWE v1.1 at 71.9% — a coding-heavy agentic evaluation that runs models on realistic software engineering tasks with constrained scaffolding.

Why Agent Arena Is Different

Agent Arena evaluates models on multi-step agentic tasks rather than single-turn responses. The benchmark set includes tasks that require planning, tool use, error recovery, and persistence across turns — the operational profile of AI agents in production deployment rather than the chatbot interface.

Standard Arena rankings measure which model a human prefers in a one-shot conversation. Agent Arena changes the question to which model a human would trust to execute a multi-step workflow. The two rankings do not always agree. A model can perform well on conversational preference tests and poorly on agent reliability, particularly if it fabricates tool calls, loses context mid-task, or hallucinates intermediate results.

Previous Agent Arena additions in October have included Claude Haiku 5.5 High (October 14), Muse Wonder Max (October 14), and a cluster of eight frontier variants added across the Search Arena on October 13. MiMo V2.6 is the first Xiaomi model added specifically to the Agent evaluation surface.

Open-Weight in the Agent Evaluation Stack

The significance of open-weight models appearing on agentic benchmarks is distinct from their appearance on capability or quality leaderboards.

Capability benchmarks (intelligence indexes, coding evals) measure what a model can do. Agentic benchmarks measure what a model does under realistic, repeated deployment conditions — including the failure cases. Open-weight models are deployable on private infrastructure, which means their Agent Arena results carry direct relevance to enterprise deployments that cannot use third-party APIs for sensitive workloads.

MiMo V2.5-Pro held a SWE-bench Verified score of 78.9% at launch in May 2026. The Flash variant of V2.6 has matched that figure (78.6%, Thinking mode). The Pro variant, at 1.02T parameters, represents one of the largest open-weight models currently in active agentic evaluation on Arena.

Trajectory

Xiaomi entered the international AI model competition in April 2026 with MiMo V2.5, positioning it against models from Alibaba, Moonshot, and Tencent rather than directly against OpenAI or Anthropic. The trajectory since has been deliberate: Code Arena rankings, the AA Intelligence Index, and now the Agent Arena leaderboard.

Each platform covers a different evaluation surface. Code Arena is task-completion on software engineering problems. The AA Intelligence Index is a composite across ten capability benchmarks. Agent Arena is multi-step agentic execution under preference evaluation. MiMo V2.6 appearing on all three within a two-week window reflects a systematic benchmarking push across every major public evaluation platform.

The Agent Arena results are not yet published. Watch the leaderboard at arena.ai/leaderboard/agent for initial rankings as human preference votes accumulate.