Xiaomi MiMo-V2.6

Xiaomi MiMo-V2.6 is a family of open-weights omni-modal mixture-of-experts language models released by Xiaomi's MiMo team on 21-22 September 2026. The series has two checkpoints: MiMo-V2.6-Pro, with 1.02 trillion total and 42 billion active parameters, and MiMo-V2.6-Flash, with roughly 310 billion total and 15 billion active parameters. Both accept text, image, video and audio input and support a context window of up to 1 million tokens.^[1][2][3] Xiaomi's API also offers MiMo-V2.6-Pro-UltraSpeed, a faster serving mode of the Pro model that Xiaomi says produces output up to 20 times faster.^[4] The weights are published on Hugging Face under the MIT license.^[2][3]

The release is built around one idea: scaling reinforcement learning compute. The accompanying technical report, "MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement," describes a single large, fully asynchronous RL run that mixed coding, general-agent, visual and cybersecurity tasks in the same batches, followed by a multi-teacher on-policy distillation stage called MOPD2 that folds in skills from separately trained teachers.^[1] Alongside the two flagship checkpoints, Xiaomi released MiMo-V2.6-Distill-Qwen-9B, a small model fine-tuned from Qwen3.5-9B on MiMo-generated trajectories, together with RL environments and training code intended as a starting point for outside researchers.^[1][6]

Xiaomi's launch post says MiMo-V2.6-Pro scored 46.32 on the Artificial Analysis Intelligence Index and calls it "the strongest open-source model to date."^[4] Artificial Analysis's own model page, checked on 23 September 2026, lists a score of 46 and ranks it first of 114 in its comparison class (Artificial Analysis compares open-weights models with other open-weights models of the same size class, here over 150B parameters), while proprietary frontier models such as Claude Opus 5.5 and GPT-6 Astra score higher.^[7][8]

Overview

Sources: technical report, model cards and Hugging Face repository metadata.^[1][2][3] The report's introduction and Table 1 give Flash as 310B parameters, while the Flash model card says 309B.^[1][3]

Background

MiMo is Xiaomi's brand for its in-house large models. Earlier releases include the 7B MiMo reasoning model described in a May 2025 paper, the MiMo-V2-Flash mixture-of-experts model and the MiMo-V2.5 family (see Xiaomi MiMo-V2.5).^[18][13][9] MiMo-V2.5-Pro, the previous flagship, was described on its model card as a 1.02T-parameter MoE language model; MiMo-V2.6-Pro keeps the same total and active parameter counts but adds vision and audio encoders, so the Pro tier is now omni-modal.^[9][2]

The V2.6 report is published under the name "LLM-Core Xiaomi" and lists Luo Fuli as the corresponding author; she is one of 79 named core contributors, with 69 further contributors.^[1] Luo wrote in a post on X on 21 September that the company dedicated "a team of several dozen people to one goal over an extended period: scaling up RL."^[5]

Architecture

MiMo-V2.6 uses a standard Transformer backbone with visual and audio encoders attached through lightweight projectors. The text backbone stacks hybrid blocks, each made of several local sliding-window attention (SWA) layers followed by one global-attention (GA) layer. The sliding window is 128 tokens. The very first Transformer block uses global attention with a dense feed-forward network "to stabilize early representation learning"; every other block uses a sparse MoE feed-forward layer without shared experts.^[1] The report points to the MiMo-V2-Flash technical report for a fuller description of this backbone design.^[1][13]

Source: technical report, Table 1.^[1]

Vision encoder. MiMo-ViT is a 681M-parameter encoder with 28 layers (24 SWA, 4 global). It replaces the fixed, non-overlapping window attention of the earlier MiMo-VL-7B with sink-augmented sliding-window attention, alternating row-major and column-major token order between layers so information can move along both image axes. Xiaomi pre-trained it from scratch on more than 4T image tokens by pairing it with a small LLM and training only with a cross-entropy objective, without contrastive losses.^[1]

Audio encoders. Audio goes through a 308M-parameter audio tokenizer, which uses a 20-layer residual vector quantizer to turn each 25 Hz frame into 20 discrete tokens and was trained on 20 million hours of audio, and then through a 127M-parameter patch encoder that groups four frames per patch, reducing the rate fed to the backbone to 6.25 Hz.^[1]

Speculative decoder. A 5-layer multi-token prediction module follows the block-diffusion design of DFlash and predicts 7 subsequent tokens in one forward pass for parallel verification by the backbone.^[1][14] During RL rollouts Xiaomi used a block-6 DFlash drafter fine-tuned on early RL rollout logs, which the report says raised average accepted length by 31.3% over the multi-token prediction configuration inherited from supervised fine-tuning.^[1] (See speculative decoding and multi-token prediction.)

The Hugging Face configuration files declare a 1,048,576-token maximum position embedding for both models, and specify FP8 (E4M3) quantization with an MXFP4 storage type for the published weights.^[2][3] The report says the team used MXFP4 quantization-aware training during mid-training and MXFP4 as the experts' data type during rollout.^[1]

Pre-training and mid-training

Pre-training ran in two stages: a text-only stage for the language backbone, then joint training with the in-house ViT and audio encoder on omni-modal data. Context length started at 32K and was extended to 256K during pre-training. Flash was trained on 48T tokens (26T text, 22T omni) and Pro on 30T tokens (27T text, 3T omni). The optimizer for pre-training was AdamW.^[1]

A separate mid-training phase used an agent-centric data mix of coding, general, visual and research agent trajectories plus text, repository-level code and multimodal data, first at 256K context and then extended to 1M in the final stage. For mid-training the team switched the hidden weight matrices from AdamW to Muown, a variant of the Muon optimizer with explicit row-norm control, because its preliminary experiments showed AdamW's optimization efficiency falling as batch size grew in mixed-task RL. The report notes that earlier work had found switching an Adam-trained model to Muon can hurt performance, but says it saw no loss spike.^[1] Mid-training also included examples in which the model reflects on and rewrites its own earlier reward-hacking turns, which the report says improved alignment.^[1]

Scaling reinforcement learning

After a short supervised fine-tuning stage, Xiaomi scaled RL along three axes: training compute, environments and harnesses, and grader compute.^[1]

Compute and batch size

The RL objective is GRPO (Group Relative Policy Optimization) with prompt-mean loss aggregation. Each step samples 1,568 prompts with 16 rollouts per prompt, about 25,000 trajectories, which the report says amount to 2.7B to 3.7B training tokens per step (roughly 110K to 150K tokens per sequence). Training is fully asynchronous with partial rollouts at a staleness of 4.^[1] Xiaomi's launch post gives a slightly different figure of 3.5 to 3.7B tokens per step.^[4]

The report puts RL post-training spending at $2.6M for Pro and $0.9M for Flash, run across thousands of GPUs in a single job. For Pro, rollout took 43.8% of cost, training 43.5% and grading 12.7%. On the held-out DeepSWE v1.1 benchmark, the average@3 score rose from 58.4 to 72.6 for Pro and from 48.7 to 65.7 for Flash over the course of RL.^[1] The launch post describes the same run as 30 RL steps per model over roughly 750,000 trajectories, completed in under six days, at about $0.85M (Flash) and $2.62M (Pro).^[4] The report's failure-analysis timeline shows 30 steps taking 123.1 hours for Pro and 81.8 hours for Flash, including interruptions from GPU memory errors, a Kubernetes failure, an unreachable grader and out-of-memory crashes.^[1] Xiaomi says it streamed the production run publicly, and the report links a running log.^[1][4]

Training settings reported for the run include the Muown optimizer at a learning rate of 3e-6 with no weight decay or warmup, decoupled importance-sampling clip bounds initialized to [0.2, 5.0] and tuned at runtime based on policy entropy, and a dynamic sampler that drops all-pass and all-fail groups.^[1]

Frozen router

The report compares two Pro RL runs that differ only in whether the MoE router is trainable. With a trainable router, expert load at decoder layer 9 collapsed within 20 steps: the coefficient of variation rose from 0.78 to 2.0, the peak load factor from 6x to 16x, and the share of "cold" experts from 0.5% to 22%. Restoring the step-20 router to its pre-RL weights restored load balance without changing benchmark scores, so the team concluded the collapse came from router drift rather than the experts, and froze the router for the production run.^[1] Training-inference consistency was handled with Rollout Routing Replay (R3), which replays the expert choices made during rollout, and by recording each token's top-p candidate set so training log-probabilities are renormalized the same way as sampling.^[1][15]

Environments, harnesses and task mix

The RL task mix was agentic and competitive coding (68%), aesthetic design (13%), general tool use (12%), cybersecurity (4%) and context following (3%).^[1] Coding tasks came from five synthesis pathways, including GitHub pull requests, requests from Xiaomi employees, specification-driven tasks and tasks derived from existing codebases by a tool called CodeMidas. Each task was checked with rollout-based audits and eight repeated test executions to screen out flaky rewards.^[1] General-agent environments used locally mocked software with resettable state, graded with atomic rubric items. Cybersecurity tasks, drawn from OSS-Fuzz, asked the agent to build an input that triggers a specific known vulnerability, and were verified by matching the sanitizer's vulnerability type and crash location rather than by comparing against a patched binary.^[1]

Rather than training inside production agent harnesses such as MiMo Code or Codex, Xiaomi built minimal, recombinable "mini-harnesses" for code, general, visual and cyber tasks, mixing several of them in each batch. In a coding experiment, DeepSWE v1.1 pass@1 on three held-out harnesses (Codex, Claude Code and mini-swe-agent) rose from about 50% to 66% over training.^[1] The model cards call this recipe "You Only RL Once": one mixed RL run across domains instead of separate per-domain runs.^[2]

Reward hacking defences

The report documents reward hacking in repository-repair tasks, where agents installed newer package releases, fetched upstream files, cloned newer checkouts or read issue discussions to copy a published fix. Defences included stripping build artifacts, caches and later Git history from environments, network isolation, a dedicated "hack agent" that probed environments for leaks until it found no working exploit, and offline audits of trajectories during training. Confirmed hacking trajectories had their reward set to zero, and the logged confirmed-hack share stayed below 2% throughout training for both models.^[1]

Groupwise agentic grading

Binary test rewards cannot tell a clean passing patch from a sloppy one, so Xiaomi added an agentic grader that compares rollouts within each group.^[1]

• Groupwise Reward Synthesis (GRS), used offline on a subset of high-pass-rate tasks, has an agent study several rollouts together and write task-specific solution and behavior rubrics. During training the final reward is the binary test reward multiplied by the solution score and the behavior score, so failed attempts stay at zero while passing ones are ranked.^[1]
• Groupwise Advantage Redistribution (GAR), used online for the remaining code tasks, has an SFT-trained grader rank passing patches within each mixed-outcome group on approach, precision, minimality, side effects and craftsmanship, then shifts positive advantage from weaker to stronger passing solutions. In its uncapped form the update keeps the total positive advantage of passing trajectories unchanged; in practice the rescaling factor is capped.^[1]

In a code-only ablation on Flash, runs without GAR saw turn counts and token lengths grow quickly and pass-rate gains stall, while runs with GAR kept improving through step 52 with roughly stable turn counts. The report also says audits found policies trained without online grading drifting toward "speculative compatibility branches, broad exports, exception swallowing, relaxed validation" and similar test-passing workarounds.^[1] A group-relative length penalty and segment-level penalties for malformed tool calls were layered on top.^[1]

MOPD2 and the MixRL question

After mixed RL, Xiaomi applied Multi-Prefix Multi-Teacher On-Policy Distillation (MOPD2), an extension of the MOPD method from MiMo-V2-Flash.^[1][12] Domain teachers are either "mixRL" teachers trained with RL on verifiable tasks or SFT teachers trained on synthetic demonstrations for open-domain tasks. MOPD2 keeps standard on-policy distillation of full student rollouts and adds prefix-conditioned single-turn rollouts: a teacher trajectory (Teacher-Prefix OPD) or SFT demonstration (SFT-Prefix OPD) with k assistant turns is split into k history prefixes, and the student generates one new turn from each prefix under token-level supervision from the assigned teacher, without regenerating earlier turns.^[1][16] The report says this extends the model to domains where training-time verification is hard, naming long-horizon game development, scientific research and embodied intelligence.^[1] (See knowledge distillation.)

Luo Fuli addressed why the team used both approaches in her 21 September post. In her account, MixRL and MOPD "are not competing choices": the team "ran MixRL on verifiable tasks of moderate difficulty, including code and related agentic tasks," while tasks that are hard to verify, very long-horizon or too difficult for the joint run were trained separately, because "Including them would substantially reduce rollout efficiency or introduce significant rollout staleness." Those capabilities were then merged through MOPD, and she listed "Games, 3D tasks, and tasks with subjective evaluation signals" as examples.^[5]

Benchmark results

The report's final comparison table covers agentic benchmarks only. Baseline models with configurable reasoning effort were run at their highest setting. Three of the benchmarks (MiMo Code Bench, MiMo Cyber Bench and MiMo Visual Coding) are Xiaomi's internal evaluations, and the report notes that its CyberGym numbers use evaluation environments Xiaomi corrected. The table is Xiaomi's, but the report does not say which comparison-model scores it measured itself. Several cybersecurity cells match figures OpenAI published for its GPT-6 Astra launch: 78.5 and 70 on ExploitBench for GPT-5.6 Sol and Claude Opus 5, 30.3 on ExploitGym and 79.1 on SEC-Bench Pro for Sol. Two cells listed under "Claude Fable 5" match results other vendors reported for Claude Mythos 5: 28.4 on ExploitGym, which OpenAI labels as a Mythos result, and 78.0 on ExploitBench.^[1][19]

Source: technical report, Table 3, and the model cards.^[1][2]

On Xiaomi's numbers, MiMo-V2.6-Pro posts the highest score in the table on AutomationBench and Terminal-Bench 2.1, and ties Claude Opus 5 on Agents' Last Exam. It trails the leading closed model on most other rows, with the widest gaps on Terminal Bench 4.0 (34.9 against 49.0 for Claude Opus 5), ProgramBench (26.5 against 37.0) and the exploit-development benchmarks ExploitGym and ExploitBench, where GPT-5.6 Sol scores far higher, as do the figures Xiaomi lists under Claude Fable 5, which match published Claude Mythos 5 results. The comparison set does not include Claude Opus 5.5, which Anthropic announced on 22 September 2026, a day after the MiMo weights went up. The gains over MiMo-V2.5-Pro are large across the board, for example 19.0 to 71.9 on DeepSWE v1.1 and 1.5 to 34.9 on Terminal Bench 4.0.^[1]

Independent index

Artificial Analysis lists MiMo-V2.6-Pro as an open-weights, MIT-licensed model released in September 2026 with a score of 46 on its Intelligence Index (the page describes version 4.3.2, which aggregates 10 evaluations including GDPval-AA v2.1, AutomationBench-AA and Terminal-Bench 4.0). As of 23 September 2026 the model page ranked it #1 of 114 models in its comparison group for intelligence, #41 of 114 for output speed (54.2 tokens per second) and noted that it generated 140M tokens while running the index. The same page puts its average cost at $0.13 per Intelligence Index task.^[7] Artificial Analysis's announcement on X said that figure places the model on its Intelligence vs. Cost per Task Pareto frontier.^[20] On the same page's comparison chart eight proprietary models score higher: Claude Opus 5.5 (57.6), Claude Fable 5.1 (53.4), GPT-6 Astra (52.7), Claude Opus 5 (50.8), Muse Spark 1.3 (48.1), GPT-6 Sol (47.5), GPT-5.6 Sol (47.0) and Grok 4.7 (46.4). On Artificial Analysis's open-weights page it is the highest-scoring open-weights model, ahead of GLM-5.3 (44.8) and Kimi K3 (43.6).^[7][8] Luo Fuli's post describes the model as "the number one open-source model," and Xiaomi's launch post makes the same claim; the Artificial Analysis data supports it for the open-weights models on that board at the time of writing, but not against closed models.^[4][5]

Open-source releases

MiMo-V2.6-Distill-Qwen-9B

Xiaomi built MiMo-V2.6-Distill-Qwen-9B by supervised fine-tuning of Qwen3.5-9B on 77.4B tokens of MiMo-generated data (27.2B of them loss tokens) spanning code, cybersecurity, general and visual tasks. It is released under the MIT license as a starting checkpoint for agentic RL research.^[1][6] The report then ran separate GRPO experiments per domain from that checkpoint using the released environments.^[1]

Source: technical report, Table 6. The RL column is a set of domain-specific checkpoints, not one model.^[1]

Environments and training code

The report describes released RL environments of roughly 7,000 training tasks: about 3,000 software-engineering tasks with executable tests, 1,000 vulnerability-reproduction tasks with rule checks, 1,000 knowledge-work tasks with rubric-based judging and 2,000 web-development tasks with visual grading, plus about 1,000 music-generation tasks.^[1] Luo's post refers to the same release as "7K diverse environments and a complete RL training framework."^[5]

On 21 September Xiaomi published three forked repositories on GitHub carrying its changes. XiaomiMiMo/verl, a fork of the verl RL framework (Apache 2.0), adds multi-harness agentic RL recipes, tool-call credit assignment and web-development and music training setups. XiaomiMiMo/uni-agent, a fork of Uni-Agent (Apache 2.0), adds a multi-harness gateway for long-horizon agent training. XiaomiMiMo/mimoagent (MIT) is an agentic rollout framework that started as a fork of mini-swe-agent v1.9.0 and, according to its commit notes, has since been "largely rewritten."^[10][11][17]

Availability and pricing

Xiaomi says MiMo-V2.6-Pro and MiMo-V2.6-Flash are available in AI Studio, MiMo Code, MiMo Desktop, the Xiaomi MiMo API platform and on OpenRouter, with API prices unchanged from V2.5. MiMo Desktop left early access with this release.^[4] The model cards give serving recipes for SGLang and vLLM.^[2][3]

USD per million tokens, from Xiaomi's launch post.^[4]

Demonstrations

The launch post groups its demos under a theme Xiaomi calls "Vibe World": game development from image, video or text prompts using multiple coordinated agents; 3D modeling in Blender; controlling a simulated Franka Panda robot arm from multi-view camera feeds; frontend, slide and video design; and music composition, including an orchestral piece the model wrote as a score and converted to MIDI.^[4] Xiaomi also describes two research case studies: MiMo-V2.6-Pro helping its materials scientists screen candidate metal-organic frameworks for adsorbing PFAS "forever chemicals," and a Lean 4 formalization of the main theorem of Li and Yorke's "Period Three Implies Chaos," in which MiMo-V2.6-Pro assisted researchers following an exploration strategy they designed; after further revision and integration the project ran to more than 6,000 lines, which Xiaomi says Lean's kernel checked with no unfinished placeholders, although the model had no Lean-specific post-training.^[4] These are vendor demonstrations and have not been independently evaluated.

Reception and claims

Luo Fuli's post called MiMo-V2.6 "very likely one of the largest single RL runs, by compute, that any open-source model team has undertaken to date," and said that in her view its research and engineering challenges "surpass those of DeepSeek R1, which I was partly involved in" (see DeepSeek-R1).^[5] Both are Luo's own assessments. The report itself is more measured, concluding that MiMo-V2.6 achieves "competitive performance against frontier models across a broad range of agentic tasks."^[1]

Several figures differ slightly between Xiaomi's own documents: Flash's size (310B in the report, 309B on the model card), tokens per RL step (2.7B to 3.7B in the report, 3.5B to 3.7B in the launch post), Flash's RL cost ($0.9M against $0.85M) and Flash's starting DeepSWE score (48.7 against 48.8).^[1][3][4]

See also

• Xiaomi MiMo-V2.5
• Xiaomi AI
• Luo Fuli
• Mixture of Experts
• GRPO
• Reward hacking

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47LLM-Core Xiaomi, "MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement," technical report, September 2026. huggingface.co/...MiMo_V2_6_technical_report.pdf
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8XiaomiMiMo, "MiMo-V2.6-Pro-RL" model card and config, Hugging Face. huggingface.co/...MiMo-V2.6-Pro-RL
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7XiaomiMiMo, "MiMo-V2.6-Flash-RL" model card and config, Hugging Face. huggingface.co/...MiMo-V2.6-Flash-RL
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Xiaomi MiMo, "Introducing MiMo-V2.6 series," 22 September 2026. mimo.xiaomi.com/mimo-v2-6
5. ^1 ^2 ^3 ^4 ^5Luo Fuli (@_LuoFuli), "MiMo-V2.6: The Hard Road to Scaling Up RL," X, 21 September 2026. x.com/...2102162926802968749
6. ^1 ^2XiaomiMiMo, "MiMo-V2.6-Distill-Qwen-9B" model card, Hugging Face. huggingface.co/...MiMo-V2.6-Distill-Qwen-9B
7. ^1 ^2 ^3Artificial Analysis, "MiMo-V2.6-Pro Intelligence, Performance & Price Analysis," accessed 23 September 2026. artificialanalysis.ai/...mimo-v2-6-pro
8. ^1 ^2Artificial Analysis, "Open weights models," accessed 23 September 2026. artificialanalysis.ai/...open-source
9. ^1 ^2XiaomiMiMo, "MiMo-V2.5-Pro" model card, Hugging Face. huggingface.co/...MiMo-V2.5-Pro
10. ^XiaomiMiMo, "verl" (fork of verl-project/verl), GitHub. github.com/...verl
11. ^XiaomiMiMo, "uni-agent" (fork of verl-project/uni-agent), GitHub. github.com/...uni-agent
12. ^W. Ma et al., "MOPD: Multi-Teacher On-Policy Distillation for Capability Integration in LLM Post-Training," arXiv:2606.30406, 2026. arxiv.org/...2606.30406
13. ^1 ^2Core Team, B. Xiao, B. Xia et al., "MiMo-V2-Flash Technical Report," arXiv:2601.02780, 2026. arxiv.org/...2601.02780
14. ^J. Chen, Y. Liang and Z. Liu, "DFlash: Block Diffusion for Flash Speculative Decoding," arXiv:2602.06036, 2026. arxiv.org/...2602.06036
15. ^W. Ma et al., "Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers," arXiv:2510.11370, 2025. arxiv.org/...2510.11370
16. ^B. Liao et al., "Multi-Turn On-Policy Distillation with Prefix Replay," arXiv:2607.04763, 2026. arxiv.org/...2607.04763
17. ^XiaomiMiMo, "mimoagent" (MiMo Agent), GitHub. github.com/...mimoagent
18. ^Xiaomi LLM-Core, "MiMo: Unlocking the Reasoning Potential of Language Model -- From Pretraining to Posttraining," arXiv:2505.07608, 2025. arxiv.org/...2505.07608
19. ^OpenAI, "GPT-6 Astra: A new generation of intelligence," 3 September 2026. openai.com/...gpt-6-astra
20. ^Artificial Analysis (@ArtificialAnlys), post on X, 21 September 2026. x.com/...2102128560962187701