An open model from Xiaomi is closing the gap with the best proprietary agents

Xiaomi unveils MiMo-V2.6 Pro and Flash, two open-weight models capable of analyzing text, images, video, and audio within a one-million-token context. The Pro version takes the lead in Artificial Analysis's open-weight ranking, without dominating all specialized tests.

A single model can inspect a software repository, analyze a screenshot, listen to a recording, review a video, and then operate external tools. That is the scope of MiMo-V2.6, a new Xiaomi family offered in two versions: Pro for demanding projects and Flash for higher-volume workloads.

Released on September 21, 2026, both models have a one-million-token context window and accept text, image, video, and audio input. “Omnimodal” needs one important qualification, however. Pro and Flash understand all four input types, but they only produce text. They do not directly generate images, video, or audio.

Their shared perception system includes a 681-million-parameter MiMo vision encoder and two audio components. Images and video frames are converted into representations that enter the model’s context, while speech and other audio signals pass through a 308-million-parameter audio tokenizer and a separate 127-million-parameter patch encoder.

The main distinction between Pro and Flash is the language backbone. MiMo-V2.6 Pro uses a Mixture-of-Experts architecture with 1.02 trillion total parameters and 42 billion activated for each token. It has 70 layers and routes each token through eight of its 384 experts.

Flash has 309 billion total parameters and activates 15 billion of them. It uses 48 layers and selects eight experts from a pool of 256. Despite its name, it is not a small local model. “Flash” refers to lower compute requirements per request, not easy deployment on an ordinary computer.

Both versions alternate local and global attention. Some layers focus on a limited window, while others examine a broader portion of the context. This design is intended to reduce memory requirements when a request contains an entire repository, several agent sessions, or a long sequence of tool results.

A five-layer speculative decoder also prepares several future response positions. It can propose seven at once, after which the main model verifies them in parallel. The goal is to accelerate output without simply reducing model capacity.

The one-million-token figure describes the maximum context capacity. Xiaomi’s API also allows as many as 128,000 output tokens. Neither number guarantees that every detail buried in a context of that size will be retrieved with equal reliability. They also say nothing about the time or total cost required to process the full window.

Xiaomi attributes much of the improvement to reinforcement learning conducted inside agent environments. Instead of running entirely separate programs for coding, tool use, visual tasks, and cybersecurity, the company says it mixed these domains within the same training phase.

One update can include 1,568 prompts with 16 trajectories generated for each prompt. The model explores several ways to complete a task, then receives feedback based on both the final result and the relative quality of those attempts.

A binary pass-or-fail signal cannot always distinguish between two successful solutions. Xiaomi therefore adds an agentic grader that compares trajectories within a group, develops task-specific criteria, and rewards stronger or more efficient approaches. This creates a self-improvement loop in which the model is partly judged against its own alternative solutions.

That process does not guarantee that the grader will always identify the genuinely best approach. As with any reward-based training process, a model may find ways to satisfy a test without solving the intended problem. Xiaomi says it uses adversarial screening, hardened environments, and cross-checking between verifiers to limit this behavior.

The mixed training phase is followed by multi-teacher distillation. Some individual decisions are trained from existing trajectories without regenerating the entire history leading up to them. Xiaomi is also releasing MiMo-V2.6-Distill-Qwen-9B, a Qwen3.5-9B derivative trained on MiMo-generated data. The smaller checkpoint is intended as a starting point for teams researching their own agent training systems.

The headline result comes from Artificial Analysis. MiMo-V2.6 Pro scores 46 on version 4.3.2 of its Intelligence Index, the highest result among open-weight models listed at launch. The independent testing organization also measured approximately 124.5 output tokens per second through Xiaomi’s API and estimated an average Intelligence Index task cost of $0.13.

That ranking does not make Pro the strongest model in every category. The index combines ten evaluations, and individual results vary considerably. Xiaomi’s own comparison table is more nuanced than its launch message might suggest.

Pro reaches 71.9% on DeepSWE v1.1, behind Claude Opus 5 at 74% and GPT-5.6 Sol at 73%. It takes a narrow lead on Terminal-Bench 2.1 with 89.9%, but falls to 34.9% on Terminal-Bench 4.0, compared with 49% for Opus 5 and 39.9% for GPT-5.6 Sol.

It leads both models on AutomationBench with 53.1%, ties Opus 5 at 31.6% on Agents’ Last Exam, and records 82% on OSWorld-Verified, slightly below the proprietary systems in the table. Xiaomi’s claim that Pro is broadly competitive across agent evaluations is reasonable, but the numbers do not show a consistent overall victory.

Cybersecurity performance is similarly uneven. Pro scores 94% on CyberGym and Flash reaches 95.1%, although Xiaomi does not provide equivalent results for Opus 5 or GPT-5.6 Sol in that row. On ExploitBench, Pro records 47.9%, well behind Opus 5 at 70% and GPT-5.6 Sol at 78.5%. Pro reaches 66.3% on SEC Bench Pro, while GPT-5.6 Sol is listed at 79.1%.

Claims around creative work and 3D also need context. MiMo-V2.6 can interpret references, write code, coordinate several agents, and inspect rendered output. Xiaomi demonstrates this by building a playable 3D world from text, images, or video. The model is not directly producing a complete 3D asset by itself. It decomposes the project, operates external tools, and reviews their results.

Pro and Flash weights are available through Hugging Face and ModelScope under