Xiaomi Released the 1.02T-Parameter MiMo-V2.6-Pro Under an MIT License and Put It in the Same Table as Claude Opus 5

Xiaomi's MiMo team published MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL on Hugging Face under an MIT license at 15:39 UTC on September 21, 2026. Pro is a sparse mixture-of-experts model with 1.02 trillion total parameters and 42 billion activated per token; Flash carries 309 billion total with 15 billion activated. Both models take a 1M-token context window and accept text, image, video, and audio in a single model. The model card ships a 17-row comparison table that places Claude Opus 5 and GPT-5.6 Sol in adjacent columns, and ASAP reads that table row by row to establish how far the phrase "leading open-weights model" actually holds.

The release consists of two checkpoints and a decision to run reinforcement learning only once

The stated design goal of the MiMo-V2.6 series is to "scale reinforcement learning toward self-improvement," and the mechanism Xiaomi names for it is You Only RL Once. Instead of separate per-domain runs, coding, general agents, visual tasks, and cybersecurity are mixed into a single reinforcement learning run, with tasks and multiple harnesses placed in the same batch so capabilities reinforce each other and strategies transfer to harnesses never seen during training.

The compute scale is on the card as well. Training uses fully asynchronous Group Relative Policy Optimization with 1,568 prompts and 16 rollouts per step, and billions of tokens per update. Two mechanisms scale the reward signal itself. Groupwise Reward Synthesis builds task-specific rubrics offline from contrasting rollouts and fuses rubric quality with test outcomes; Groupwise Advantage Redistribution ranks passing trajectories online and shifts advantage toward higher-quality solutions. Both target the same weakness: a binary pass or fail signal cannot rank the solutions that already pass.

The Pro backbone is 70 layers, of which 60 use sliding window attention and 10 use global attention. Hidden size is 6,144, there are 384 routed experts with 8 activated per token, and no shared experts. A 681M-parameter MiMo ViT handles vision; a 308M AudioTokenizer and a 127M patch encoder handle audio. A five-layer speculative decoder predicts seven subsequent tokens per forward pass. Flash is the same design reduced to 48 layers, hidden size 4,096, and 256 experts.

Counted row by row, MiMo-V2.6-Pro beats Claude Opus 5 in four of fourteen cells

MiMo-V2.6-Pro meets Claude Opus 5 head-to-head in 14 rows of Xiaomi's table, and leads or ties in four of them. Pro takes AutomationBench v1.0.6 at 53.1 against 50.3, Terminal Bench 2.1 at 89.9 against 89.1, and MiMo VisualCoding at 72.3 against 70.0, and ties Agents' Last Exam at 31.6.

Claude Opus 5 takes the remaining ten. DeepSWE v1.1 is 71.9 against 74.0, ProgramBench 26.5 against 37.0, MiMo Code Bench 63.2 against 68.6, Toolathlon-Verified 76.9 against 80.6, GDPval-AA 2.1 1,673 against 1,708, Terminal Bench 4.0 34.9 against 49.0, OSWorld-Verified 82.0 against 83.4, and JobBench 62.0 against 65.7. The cybersecurity gaps run wider: ExploitGym is 17.8 against 22.1 and ExploitBench is 47.9 against 70.0.

Measured against the previous generation instead, the picture inverts. MiMo-V2.5-Pro scored 19.0 on DeepSWE v1.1 and this release scores 71.9, a gain of 52.9 points. Terminal Bench 4.0 moves from 1.5 to 34.9, JobBench from 25.0 to 62.0, Toolathlon-Verified from 49.1 to 76.9, and GDPval-AA 2.1 from 1,107 to 1,673. MiMo Cyber Bench goes from 0.0 to 80.2. One generation ago this line had effectively no score on agentic tasks; it now appears in the same table as frontier models and loses ten rows.

Terminal Bench 2.1 at 89.9 and Terminal Bench 4.0 at 34.9 come from the same model

ASAP's reading starts here: the 55.0-point spread between two versions of the same benchmark on one model is the most informative fact in the table. MiMo-V2.6-Pro scores 89.9 on Terminal Bench 2.1, ahead of Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8, then scores 34.9 on 4.0, behind Claude Opus 5 at 49.0 and GPT-5.6 Sol at 39.9.

Across those same two rows, Claude Opus 5 drops 40.1 points and GPT-5.6 Sol drops 48.9. MiMo's 55.0 is the largest of the three. A shape that runs high on the older version and low on the newer one indicates either that the older harness is near saturation or that the model is fitted more tightly to the published task distribution. Either way, the row worth carrying into a deployment decision is 34.9, not 89.9. The newer version is harder because tasks run long and intermediate results have to be checked by the model itself, and real work resembles that side.

The composition of the table deserves attention too. Three of the 17 rows, MiMo Code Bench, MiMo Cyber Bench, and MiMo VisualCoding, are benchmarks Xiaomi built. One of Pro's four wins over Claude Opus 5, MiMo VisualCoding, sits in that group. Publishing in-house benchmarks is a transparency gain, but those rows do not carry the same weight when counting wins. Excluding all three, the advantage over Claude Opus 5 narrows to two cells plus one tie.

In the cybersecurity rows, Flash outscores Pro in one cell

The five cybersecurity rows are the most contested part of this release. Flash takes CyberGym at 95.1 against Pro's 94.0, and the order reverses on MiMo Cyber Bench at 80.2 for Pro and 77.2 for Flash. The remaining three separate sharply: ExploitGym is 17.8 for Pro and 6.0 for Flash, ExploitBench is 47.9 and 25.3, and SEC Bench Pro is 66.3 and 47.5.

What that distribution says is straightforward. On tasks closer to automated vulnerability discovery, a model activating 15 billion parameters keeps pace with one activating 42 billion; as the task shifts toward completing a working exploit, the gap roughly doubles. And on exactly those harder rows, both MiMo models trail every closed model in the table. ExploitBench reads 78.5 for GPT-5.6 Sol, 78.0 for Claude Fable 5, 70.0 for Claude Opus 5, and 47.9 for Pro.

From a misuse standpoint the decisive variable is the license, not the ranking. A model that names cybersecurity as an explicit reinforcement learning domain has been released weights-and-all under MIT, which permits commercial use, redistribution, and modification with no use restrictions attached. Set against the common practice of shipping open weights under bespoke licenses carrying acceptable-use clauses, the same capability now arrives with far thinner contractual constraints. Both facts hold together: the published scores still trail the frontier, and those scores are the kind that fine-tuning moves.

The real barrier sits between an MIT license and 573.5 gigabytes

The MiMo-V2.6-Pro-RL repository holds 573.5 gigabytes of weights across 132 safetensors files. The repository carries fp8 and 8-bit tags, so that figure is already the compressed footprint. This is the point where a free license and a runnable model separate.

The SGLang command the model card recommends uses tensor parallelism of 16, data parallelism of 2, expert parallelism of 16, and two nodes. The vLLM example uses tensor parallelism of 8. Xiaomi's own baseline configuration is therefore multi-node. Filling the 1M-token context window adds KV cache on top of that, which means the 1M in the spec sheet and the length usable in a single request are not the same number.

For teams outside the largest infrastructure operators, the release splits in two directions. Where network separation and data-residency rules apply, in finance, public sector, and manufacturing, an MIT-licensed model with a 1M-token context window plus audio and video input is a short list of one or two. Where 573.5 gigabytes across two nodes is not available, the practical meaning of this release is third-party hosting rather than self-hosting, and the remaining advantage is price and capability rather than the license. The model card lists the Xiaomi MiMo Open Platform API and OpenRouter among the access paths. VentureBeat reported Pro pricing at $0.435 per million input tokens and $0.87 per million output tokens, and Flash at $0.14 and $0.28, figures the model card itself does not state.

Every cell in the table was measured in Xiaomi's own environment

Three verification limits remain on the MiMo-V2.6 table published on September 21, 2026, and each one narrows what the 17 rows establish. First, the Claude Opus 5, GPT-5.6 Sol, and Claude Fable 5 scores in the comparison table are values Xiaomi measured in its own environment, not third-party reproductions, and the card does not state per-row inference settings or compute budgets for the competing models.

Second, a number of cells are empty. The CyberGym and MiMo Cyber Bench rows carry no competitor scores at all, and the Flash cell on GDPval-AA 2.1 and the V2.5 cell on OSWorld-Verified are blank as well. A blank means neither ahead nor behind, but it does block any attempt to compress the table into a single sentence. Third, at the time of checking, the Hugging Face download counter reads 0 while likes read 291 for Pro and 292 for Flash. Those are the numbers of a repository the community has not yet run at scale, which means no external reproduction exists outside this table.

One sentence summarizes the release. Xiaomi lifted agentic benchmark scores from near zero to just under the frontier in a single generation, published the result under MIT, and beat Claude Opus 5 in two cells once its own benchmarks are set aside. The distance between "leading among open weights" and "caught up to the frontier" is exactly those ten rows.

Source: Hugging Face model cards for XiaomiMiMo/MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL, Hugging Face model API metadata, VentureBeat report of September 21, 2026