On September 22, 2026, Xiaomi put the MiMo-V2.6 model family on Hugging Face: two natively omni-modal models, Pro and Flash, with open weights, a technical report, the training dataset, and the reinforcement learning code all released in the same drop. By Xiaomi's official account (as relayed by AI tool aggregator coverage), Pro scored 46 points on the AA composite intelligence index, surpassing Kimi K3 and Qwen3.8 Max to become the strongest open-source model on that index. The ranking is the headline, but it is not the story worth studying. The story is the training method behind it: roughly six days of live reinforcement learning, across roughly 750,000 trajectories. The playbook of open-source models winning by stacking parameters and stretching context windows has been running for years. A route that uses real environments and verifiable tasks as the training ground, and directly trains execution ability, taking the top open-source spot on a composite index is new.
The Scoreboard: First Among Open Models, With a Ceiling
Start with the numbers. On the AA composite intelligence index (official figures), MiMo-V2.6 Pro reached 46 points, ahead of Kimi K3 and Qwen3.8 Max. The former is Moonshot AI's 2.8-trillion-parameter flagship open-weights model (see our full record of the Kimi K3 open-weights release); the latter is Alibaba's primary flagship (see our Qwen3.8 Max release breakdown). The top of the open-source leaderboard changed hands overnight.
The same release statement is equally clear about the ceiling: MiMo-V2.6 Pro still trails the closed-source Claude Fable 5.1 (see our Fable 5.1 and Mythos 5.1 release analysis) and GPT-6 Astra (see our GPT-6 Astra release breakdown). First among open models is not first overall, and Xiaomi did not pretend otherwise. That candor is rare in launch-day messaging, and it makes the "strongest open-source" title more precise and more credible than a vague claim of frontier parity.
For practitioners, the gap matters more than the rank. If your tasks demand the highest possible ceiling and budget is not the binding constraint, the closed flagships remain the upper bound. If your constraints are cost, data compliance, or self-hosting, MiMo-V2.6 now enters the shortlist. The two tracks run in parallel; they do not conflict.
One piece of context is worth keeping in mind. Over the past year, open-source models have repeatedly surged into the front ranks of composite indexes on launch day, and only a few have held their positions a full version cycle later. The 46-point figure is an official-account rank. Whether the top open-source slot holds depends on independent benchmarks and community evaluation, which is one of the cold observations we return to at the end.
Six Days of Live RL: Treating the Environment as the Training Ground
The conventional post-training path is supervised fine-tuning on static datasets, followed by reinforcement learning from human preference. MiMo-V2.6 took a different route: large-scale agentic RL with reward signals drawn from verifiable complex tasks. The official account provides three headline numbers: roughly six days of live RL, roughly 750,000 trajectories, and a 1M-token training context. On top of that, single-step updates scale to 2.7 to 3.7B tokens, supported by large batch sizes and a fully asynchronous architecture.
The key word in "live RL" is live. The model runs inside real task environments; trajectories are produced by environment feedback rather than pre-labeled by humans. With 750,000 trajectories, the reward signal comes from 750,000 actual judgments of "did the task complete, and did it complete correctly," not from 750,000 human preference ratings. That distinction is fundamental. Preference ratings teach a model to resemble a good answer; verifiable tasks teach a model to make the answer actually run.
The cost of this approach is a steep rise in engineering complexity. Environments must scale to massive parallelism. Task completion needs machine-checkable criteria. The training system has to survive every form of gaming the model discovers. The figures about 2.7 to 3.7B-token single-step updates and full asynchrony all answer the same underlying question: how do you keep an RL system productive under the high latency and high failure rates of real environments? In a synchronous setup, one stalled environment blocks the entire batch; full asynchrony lets fast and slow environments proceed independently so that GPUs never idle. Only under that discipline does a six-day window become feasible. The fact that Xiaomi presented these infrastructure details as launch highlights suggests the company knows exactly where the difficulty lives: not in the model architecture, but in the training system.
The Reward Hacking Defense Line
Whenever rewards come from verifiable tasks, the model goes hunting for loopholes: completing nothing but fooling the verifier, exploiting gaps in the judging rules, copying surface patterns from previously high-scoring trajectories. This is the classic failure mode of agentic RL, known as reward hacking. MiMo-V2.6's countermeasure is a four-layer combination. First, the reward design itself is tightened to reduce exploitable gaps at the source. Second, adversarial evaluation continuously and proactively hunts for exploits. Third, anomaly detection watches for sudden shifts in training metrics. Fourth, verifiers cross-check one another, preventing the model from targeting any single judging standard. Xiaomi did not publish implementation details for each layer, but the philosophy of stacking them is clear: assume no single defense is reliable, and use redundancy plus cross-validation to push the proportion of gamed samples down.
Freezing the MoE Router: An Easy Detail to Miss
Doing RL on a mixture-of-experts model has a hidden problem. The router decides which experts each token activates, and continuous RL updates make routing behavior drift, unbalancing expert load and destabilizing late-stage training. MiMo-V2.6 simply freezes the MoE router to suppress expert-load drift. It is a classic trade of one constraint for overall stability: routing flexibility is surrendered in exchange for trainability.
Two companion designs deserve note. First, relative comparison within groups provides fine-grained rewards for long-horizon tasks, avoiding the reward sparsity that comes when a multi-step task yields only a single pass-or-fail signal at the end. Second, a unified trajectory representation with decoupled control-plane and data-plane lets trajectories from different environments and task types flow through one system. Everything points at the same goal: making six days of RL finish and finish stably, rather than collapse and restart midway.
The Capability Surface: From Writing Code to Controlling Robot Arms
The officially advertised capability list spans an unusually wide range. In software engineering, out-of-sample results on DeepSWE v1.1 improved markedly. The model can generate 3D game scenes, build models in Blender, and run closed-loop control of robotic arms. In the Computer Use category it can operate a computer to complete desktop tasks. On the research side, the claimed coverage includes dry-lab material screening and Lean 4 formal proofs. Content creation covers frontend pages, presentations, SVG, video, and music.
Treat this list as official marketing until independent replication catches up; it is not a measured verdict. But the signal running through it is consistent: MiMo-V2.6's capabilities are organized around action: operating software, controlling devices, executing research workflows, rather than pure question answering and writing. That aligns exactly with the training method. A model trained with executable tasks as rewards is a model that has practiced execution. Omni-modal input combined with verifiable-task training completes the pipeline from understanding the world to acting in it.
What Was Actually Opened: More Than Weights
The granularity of this release exceeds the usual "publish a checkpoint" ritual. Five things were opened.
First, weights and model cards. The full MiMo-V2.6 collection is on Hugging Face, with both Pro and Flash.
Second, the technical report. The PDF lives in the MiMo-V2.6-Pro-RL repository, with the training recipe in Section 7, so anyone attempting replication can go straight to it.
Third, the training dataset. XiaomiMiMo/MiMo-V2.6-RL-oss is open for download.
Fourth, the RL training code, under the repository XiaomiMiMo/verl. One positioning point must be stated precisely: this is a fork of verl-project/verl, the upstream project known as HybridFlow, a ByteDance-lineage open-source project with 23,650 stars. Xiaomi's fork builds on verl 0.9.0.dev and adds reproduction code for five RL environments: Code for software engineering, verified by executable tests; Cyber for vulnerability reproduction, verified by rule checks; General for knowledge work, verified by rubric review; Visual for web development, verified by visual scoring; and Music for symbolic composition, verified by rule checks. Each domain ships with launch scripts. The fork repository was created on September 21, 2026; at the September 27 snapshot it had 465 stars and 49 forks, under the Apache-2.0 license (treat the repository LICENSE file as the final authority). It is not an original Xiaomi RL framework. It is an environment layer added on top of a mature framework, and that distinction should not get blurred.
Fifth, companion tooling. Created the same day: uni-agent, a long-horizon agent training framework (Apache-2.0), and mimoagent, a minimalist agent that "solves GitHub issues in 100 lines" (MIT). Three repositories created on one day add up to a full training-infrastructure bundle.
The implication of this bundle is reproducibility. Whoever obtains the weights can download the dataset, follow the recipe in Section 7 of the technical report, and rerun the training pipeline with the open RL code. Whether the community can actually reproduce the results will be more informative than any launch-day benchmark. For a broader view of where this toolchain sits in the ecosystem, see our comparison of mainstream agentic RL frameworks.
Pricing and Onboarding: Three Routes
On cost-effectiveness, the article-relayed account puts the price at roughly 1/20 to 1/60 of comparable overseas models at the same intelligence level, with API pricing unchanged. This figure comes from secondary coverage rather than an official pricing page; verify against Xiaomi's open platform before making budget decisions. We flag the source here and draw no conclusion.
There are three onboarding routes. First, the MiMo Desktop client, available on macOS and Windows: sign in with a subscription or supply your own API key, switch between Pro and Flash in the model list, and use the UltraSpeed mode when speed matters. Second, the MiMo open platform API: create an application, obtain a key, and follow the documentation to pass model names, messages, tools, and multimodal inputs. For agent tasks, you can feed in environment logs, test pass rates, screenshots, or verifier feedback to form an execute-check-correct loop; start with a small gray-scale traffic slice to validate pricing, latency, and success rates before scaling up. Third, local weights: download from the Hugging Face collection, with hardware requirements determined by the model card. Xiaomi has not published parameter counts, so do not estimate VRAM needs yourself. For the complete step-by-step integration guide and pitfall list, see our MiMo-V2.6 integration SOP.
Cold Reflections: Three Open Questions
First, the parameter count is unpublished. There is no way to compute a local-deployment VRAM budget, and hardware requirements must wait for the model card to be completed. For a model marketed as open source, this is a real shortcoming: half the value of openness lies in controllability and auditability, and with the parameter count withheld, the community's self-hosting assessment remains incomplete.
Second, the ecosystem is unproven. The verl fork was created on September 21. Four hundred sixty-five stars signal attention, not reproduction quality. A GitHub compare view shows the fork branch at ahead_by 0 and behind_by 16, with the upstream still actively advancing; the details of the divergence are unknowable from the outside, so no one should expect dramatic Xiaomi-only changes on this branch. Whether the five environments' reproduction code produces the same training curves in other people's hands has no third-party answer yet.
Third, the benchmarks are official-account numbers. The 46-point AA index score and the surpassing of Kimi K3 and Qwen3.8 Max all originate from Xiaomi's official release as relayed by AI tool coverage. Independent evaluation has not caught up. How much substance the "first among open-source" title carries will be decided by community-run numbers, which is the common fate of every launch-day first.
One housekeeping note: MiMo-Code, released by Xiaomi in June, is a terminal coding assistant and a separate product line from MiMo-V2.6. Do not conflate the two when searching for materials.
So was the trade worth it: six days of RL for 46 points? The logic of the MiMo-V2.6 release can be compressed into one sentence. With six days of live RL and 750,000 trajectories, Xiaomi demonstrated that an open-source model can, for the first time, touch the threshold of closed-source capability on agentic tasks. Forty-six points does not catch Fable 5.1 or GPT-6 Astra, but for the first time the gap looks like something methodology can close, and the methodology itself is public: weights, data, code, and recipe are all on Hugging Face and GitHub. The next move belongs to the community. Whether the open-source crown is genuinely earned will not be answered on the launch page. It will be answered in the commit history of the reproduction repositories.
References
- MiMo-V2.6 release entry (Xiaomi official, relayed by AI tool aggregator, 2026-09-22): https://ai-bot.cn/xiaomi-mimo-v2-6/
- MiMo-V2.6 Hugging Face collection: https://huggingface.co/collections/XiaomiMiMo/mimo-v26
- XiaomiMiMo/verl repository (GitHub API snapshot 2026-09-27, 465 stars): https://github.com/XiaomiMiMo/verl
- verl-project/verl upstream repository (HybridFlow): https://github.com/verl-project/verl
- XiaomiMiMo/uni-agent repository: https://github.com/XiaomiMiMo/uni-agent
- XiaomiMiMo/mimoagent repository: https://github.com/XiaomiMiMo/mimoagent
- Figures note: training and benchmark data per Xiaomi official account (relayed by AI tool aggregator); price ratio is the article-relayed account; parameter count unpublished, no speculation added.