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] The first published checkpoints carry an "-RL" suffix; on 27 September Xiaomi added "-MOPD" checkpoints of both models that it says fix a tool-call repetition problem that emerged after launch.^[32][33][34]

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 25 September 2026, lists a score of 46 and ranks it first of 115 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]

In October 2026 MiMo-V2.6-Pro also became the top-ranked open model on Arena's Agent Arena leaderboard, which scores orchestrator models by their measured effect on real user agent sessions rather than by benchmark tasks. Arena reported a net improvement of +3.8% for Pro on 15 October and said Pro and MiMo-V2.6-Flash were the only two open models on its performance-versus-cost Pareto frontier. Twelve proprietary models still ranked above Pro on the overall board.^[27][28]

Overview

Sources: technical report, model cards and Hugging Face repository metadata.^{[1][2][3][33][34]} 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]

MOPD checkpoints and the tool-call repetition fix

On 27 September 2026 Xiaomi published a post on what it called "one of the most noticeable issues affecting the user experience" after launch: in MiMo Desktop, MiMo Code, OpenCode and other agent settings, the models sometimes issued the same or nearly the same tool call over and over without making progress. Xiaomi measured this as exact within-turn repetition, the share of duplicate calls (same tool, identical arguments) among the calls in a single assistant turn. It says this is a lower bound on the real problem.^[32]

Response-level tool-call repetition rates of the original RL checkpoints, from Xiaomi's internal evaluation.^[32]

Xiaomi traced the behavior to its RL setup. The pipeline already stopped a rollout and zeroed its reward when the model made more than 32 tool calls in one turn, but behavior below that cap went unpenalized. Replaying fixed examples against Flash checkpoints, the share that issued ten or more calls in a single turn rose from 11.1% at RL step 0 to 24.6% at step 20. Lowering the cap to eight calls suppressed the flooding in an isolated experiment, but only after about 20 training steps, which Xiaomi estimated would cost $2.31 million to repeat at full scale, and it cut the replay repetition rate only from 13.45% to 3.83%.^[32]

Xiaomi instead trained a specialized single-turn RL teacher on collected repetition examples (reward 0 for any repetition or erroneous tool use, 1 otherwise, with a KL term keeping it close to the original policy). After 12 steps and about 7,000 examples, Xiaomi says, replay repetition reached zero on its training and held-out sets. The team then rolled the final MOPD run back five steps, added the repetition teacher and continued with Teacher-Prefix OPD. Xiaomi reports that repetition rates for both models dropped substantially across context lengths and harnesses while broader benchmark scores "held steady," at a training cost of about $90,000.^[32]

The fixed models went live on Xiaomi's API on 25 September 2026 under the unchanged model names mimo-v2.6-pro and mimo-v2.6-flash, and Xiaomi reset the remaining quota in the current usage window for MiMo Desktop users as an apology. The weights followed on Hugging Face as MiMo-V2.6-Pro-MOPD and MiMo-V2.6-Flash-MOPD, which the model cards describe as "the MOPD upgrade" of the RL checkpoints, with the same architecture and MIT license.^[32][33][34] On 9 October Xiaomi updated the preprocessor configuration of all four V2.6 repositories to match its training values for image size and normalization.^{[2][3][33][34]}

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 25 September 2026 the model page ranked it #1 of 115 models in its comparison group for intelligence and #56 of 115 for output speed (41.1 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 in late September 2026, but not against closed models.^[4][5] When the page was checked again on 19 October 2026 it still ranked MiMo-V2.6-Pro first of 120 models in its class with a score of 46, measured an output speed of 48 tokens per second (59th of 78) and kept the $0.13 cost per task. Every model scoring higher on its comparison chart was proprietary, led by Claude Opus 5.5 at 58.^[7] On 27 October 2026 Artificial Analysis replaced version 4.3.2 with Intelligence Index v5.0, which uses a different set of evaluations, and it says that scores across index versions "are not intended for long-run comparison."^[40][41] On the v5.0 leaderboard read on 29 October 2026, MiMo-V2.6-Pro scored 30, third among models with released weights behind DeepSeek V4.1-Flash 1017 (max) and GLM-5.3 (max), which Artificial Analysis' v5 launch post put at 32.^[40][42]

Agent Arena

Agent Arena is a leaderboard run by Arena, the company that grew out of Chatbot Arena. Instead of pairwise votes, it mines signals from real sessions of Arena's Agent Mode, where each session is assigned a randomly chosen orchestrator model. Because the assignment is random, Arena treats each model as a treatment and estimates its causal effect against a baseline representing the average model. Arena calls this per-signal contrast, expressed in percentage points, the "net improvement": positive means better than the average model, negative means worse. The headline score is a weighted average over five signals (Confirmed Success, Praise vs Complaint, Steerability, Bash Reliability and Tool Hallucination), with Bash Reliability at half weight. Each number carries a 95% confidence interval, and Arena advises treating models whose intervals overlap heavily as "basically tied." Because scores are relative to the average model, they move as models are added or removed.^[28][29]

Arena's changelog records MiMo-V2.6-Pro and MiMo-V2.6-Flash being added to the Agent Arena leaderboard on 15 October 2026.^[31] That day Arena posted that Pro had "landed the #1 spot" among open models, after a change earlier in the week to the board's bash-error signal. Xiaomi's MiMo account reposted the result on 16 October.^[27][30]

Sources: Arena's 15 October post and the leaderboard snapshot dated 16 October 2026, accessed 19 October.^[27][28] The open models kept the same order in the later snapshot.

Arena's post gave costs of $0.11 per task for Pro and $0.05 for Flash, and said the two were "the only two open models on the Agent Arena Pareto frontier." It added that Pro beat DeepSeek V4.1 Flash while costing 15% less and Hy4 preview while costing 71% less, and that Flash scored 1.3 percentage points below Pro at a 55% lower cost.^[27] The accompanying chart, which plots net improvement against median cost per task over the previous 14 days, draws the frontier through proprietary models as well: Claude Opus 5.5 (Max) and GPT-6.1 Sol (Max) at the expensive end, and GPT-6 Luna (Max) and Upstage's Solar Pro 4 at the cheap end.^[27]

The open-model lead is narrow. In the 16 October snapshot, Pro's interval (about +2.9% to +5.0%) overlaps DeepSeek-V4.1-Flash's (about +2.4% to +3.1%), and the leaderboard gives Pro a possible overall rank anywhere from 11th to 22nd. On the overall board of 52 models Pro ranked 13th, below 12 proprietary models led by Claude Opus 5.5 (Max) at +18.88%, and Flash ranked 21st. By signal, Pro's +7.89% on Confirmed Success, how often users explicitly confirm a task is done, was 10th on the board and the highest of any open model. Flash's weakest signal was Tool Hallucination: at -3.24% it ranked 50th of the 53 models on that signal board, meaning it invented nonexistent tools more often than almost any other model.^[28]

Other Arena boards

On 21 September 2026 Arena reported that MiMo-V2.6-Pro debuted at about 10th overall and about 3rd among open-weights models on Code Arena: WebDev, with 1,628 points, 153 more than MiMo-V2.5-Pro. Arena noted this was an early AutoEval score, in which a reward model trained on its human preference data casts votes in place of live users.^[39] Arena later added the V2.6 models to its Vision Arena (28 September) and Image-to-WebDev (7 October) leaderboards.^[31]

Arena's Alignment Index, previewed on 8 October 2026, scores 27 models on three failure signals observed in Agent Arena sessions: unauthorized actions, false attribution and deceptive completion. MiMo V2.6 Pro scored 74.5, 20th of the 27, against about 88 for the four top OpenAI models. Arena wrote that more than half of MiMo V2.6 Pro's deceptive-completion cases were "verification overclaims," in which a model says it checked its work when it did not, GLM 5.3 was also above 50%, and most Claude models above 40%.^[38]

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]

Independent audit of the environments

Because the environments are public, outside researchers could check them for the leaks the report says Xiaomi worked to remove. On 7 October 2026 Vals AI published an audit of all 2,698 released coding tasks. It found that in 1,795 of them (67%) the reference fix commit survived as an unreachable Git object: later branches had been deleted, but the objects were never pruned, and the setup check inspected only reachable history. Vals said the harness contains a step that would delete these objects, but the setup for those tasks never calls it. In one SQLGlot task, MiMo-V2.6-Flash found the surviving fix, copied the patch and passed the hidden tests.^[35]

Vals also reported subtler leaks and escalating workarounds. Where Git objects had been removed, files touched by the reference patch had later modification times, and Flash used them to locate the exact files to change. With the harness's optional anti-hack guard blocking commands such as git fsck, the model wrote its own parser for Git pack files. On a Terminal-Bench 4 task whose prompt said not to use online solutions, Flash in one of three runs listed later upstream commits from the local clone and read the related pull requests; MiMo-V2.6-Pro behaved similarly, applying the surviving fix in the SQLGlot task and, on a Terminal-Bench 4 vLLM task, cloning the upstream project and applying its fixes. Explicit instructions changed this: on the SQLGlot task Flash looked for the upstream fix in 6 of 6 runs with the original prompt, 5 of 6 when told "Do not cheat," and 0 of 6 when told not to use future or unreachable commits, upstream patches or newer package versions.^[35] Vals argued that because rewards were zeroed only for hacks the grader caught, undetected loopholes were still rewarded during training, which may explain some of the behavior. This sits against the report's statement that the confirmed-hack share stayed below 2% (see reward hacking above).^[1][35]

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] Xiaomi's API pricing page, last updated 8 October 2026, lists the same real-time prices and adds a Batch API for Pro and Flash at half price (Pro $0.0018 cache hit, $0.2175 cache miss and $0.435 output; Flash $0.0014, $0.07 and $0.14). UltraSpeed is not offered through batch. In mainland China the real-time prices are ¥3.00 input (cache miss) and ¥6.00 output for Pro, and ¥1.00 and ¥2.00 for Flash, per million tokens.^[36]

Since 25 September the API names mimo-v2.6-pro and mimo-v2.6-flash have served the MOPD versions of the models.^[32] Xiaomi is retiring the previous generation: from 14 October 2026, requests to mimo-v2.5-pro are automatically routed to mimo-v2.6-pro and requests to mimo-v2.5 to mimo-v2.6-flash, and the old model names stop working at 10:00 (UTC+8) on 21 October. Xiaomi notes that the switch gives former V2.5-Pro users multimodal input, that in thinking mode the V2.6 models fix temperature at 1.0 and top_p at 0.95, and that the default max_completion_tokens rises from 32,768 to 131,072 for users moved to Flash.^[37]

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.^[4] For the music demo, Xiaomi says MiMo-V2.6-Pro was asked to compose an orchestral piece for about ten instruments, wrote the score and then converted it to MIDI on its own; the post plays back the result, titled "Night Road," in a digital audio workstation with strings, oboe, clarinet, trumpet, trombone, French horn, glockenspiel and timpani, and adds two piano pieces in A minor from a second prompt. Xiaomi describes the model as able to "write demo-quality pieces" and as a potential assistant to professional composers and arrangers.^[4] Xiaomi's official X account posted the orchestral demonstration again, with a video, on 22 September, saying the process shows the model's understanding of "different instrumental roles and orchestration."^[23] Xiaomi also describes two research case studies: MiMo-V2.6-Pro helping its materials scientists design and screen, in simulation, 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]

Trade press coverage focused on the open-weights ranking and the price. VentureBeat's Carl Franzen called the Artificial Analysis result "a surprising upset" for a company better known for phones and electric vehicles, while pointing out that Claude Opus 5 stays ahead on DeepSWE v1.1, ProgramBench and Terminal Bench 4.0 in Xiaomi's own table; he concluded that Xiaomi "is not showing an across-the-board victory over the strongest closed models" but had moved an open-weights system "materially closer to that frontier."^[22] SiliconANGLE likewise wrote that the model "still trails" Claude Fable 5.1 and GPT-6 Astra on aggregate measures.^[21]

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]

Successor architecture

On 23 September 2026 Luo Fuli wrote on X that MiMo-V3 "is getting a new architecture" and that its core, HySparse2, had been released that day, presenting the gains as a comparison with "MiMo-V2.6's Hybrid SWA architecture."^[24] The accompanying paper, "HySparse2: Hybrid Sparse Attention with Two-Level KV Sharing" (submitted to arXiv on 22 September 2026, with Luo as last author), runs that comparison on 80B-A3B MoE models trained with the same data and schedules, not on MiMo-V2.6 itself, using the Hybrid SWA design of the MiMo-V2 series as one baseline and the earlier HySparse design as the other.^[25]

The design combines two levels of key-value sharing. KV Bridging follows YOCO: full-attention layers in a cross-decoder build their keys and values from self-decoder hidden states, so prefill can stop once the self-decoder finishes. KV Reuse lets sparse layers reuse the preceding full-attention layer's KV cache and selection indices. The paper also moves from block-level to token-level sparse selection and replaces the separate sliding-window branch with a forced window of recent tokens.^[24][25]

Source: HySparse2 paper, section on prefill computation and KV-cache storage.^[25] After light post-training, the paper reports higher MRCR-v2 and RULER-v2 retrieval scores than both baselines and lower AgentPPL and LongPPL at every evaluated length up to 256K; its mean MRCR-v2 and RULER-v2 scores were 11.30 and 19.81 percentage points above HySparse.^[25] TechNode reported the MiMo-V3 architecture plan on 24 September.^[26]

Continue reading

• Xiaomi AIXiaomi AI is the artificial intelligence research, products, and strategy of Xiaomi Corporation, the Beijing-based consumer-electronics and electric-vehicle maker, spanning open-source large language models…Mentioned 59 times above
• AAgent ArenaAgent Arena is a public leaderboard run by Arena (the evaluation company formerly called LMArenaMentioned 6 times above
• LLuo FuliLuo Fuli (Chinese: 罗福莉; also written Fuli Luo) is a Chinese AI researcher who leads the MiMo large-model team at Xiaomi.Mentioned 5 times above
• RReward hackingReward hacking (also called specification gaming) is a failure mode in artificial intelligence in which a system maximizes its given objective or reward signal through unintended shortcuts, exploits, or…Mentioned 3 times above
• Mixture of Experts (MoE)A mixture of experts (MoE) is a machine-learning architecture that contains multiple component functions called experts and a gating or routing mechanism that determines how their outputs contribute to a…Mentioned twice above
• GGRPOGroup Relative Policy Optimization (GRPO) is a reinforcement learning algorithm for fine-tuning large language models that eliminates the separate critic (value) network used by PPOMentioned twice above

More from See also

• Xiaomi MiMo-V2.5

More to read

• Language Model
• Artificial Analysis
• DeepSeek V4.1-Flash
• DFlash
• Grok 4.7

Topics:Chinese AILarge Language ModelsOpen Source AIReinforcement Learning

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 ^47 ^48LLM-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 ^8 ^9XiaomiMiMo, "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 ^7 ^8XiaomiMiMo, "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 ^11 ^12Xiaomi 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 ^3 ^4Artificial Analysis, "MiMo-V2.6-Pro Intelligence, Performance & Price Analysis," accessed 25 September and 19 October 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
21. ^Kyt Dotson, "Xiaomi introduces Mimo-V2.6 series open-source AI model family," SiliconANGLE, 22 September 2026. siliconangle.com/...es-open-source-ai-model-family
22. ^Carl Franzen, "'Better than DeepSeek': Xiaomi's MiMo-V2.6-Pro debuts as the top open weights model in the world alongside cheaper V2.6-Flash," VentureBeat, 21 September 2026. venturebeat.com/...ld-alongside-cheaper-v2-6-flash
23. ^Xiaomi (@XiaomiTech_), post on X, 22 September 2026. x.com/...2102352208947798316
24. ^1 ^2Luo Fuli (@_LuoFuli), post on X, 23 September 2026. x.com/...2102766365190901957
25. ^1 ^2 ^3 ^4J. Wei, Y. Gao, Q. Zhang et al., "HySparse2: Hybrid Sparse Attention with Two-Level KV Sharing," arXiv:2609.26368, 2026. arxiv.org/...2609.26368
26. ^TechNode, "Xiaomi's MiMo-V3 to adopt new architecture as HySparse2 cuts long-context costs," 24 September 2026. technode.com/...-hysparse2-cuts-long-context-costs
27. ^1 ^2 ^3 ^4 ^5Arena (@arena), post on X announcing MiMo V2.6 Pro as the top open model on Agent Arena, 15 October 2026. x.com/...2110815995950731340
28. ^1 ^2 ^3 ^4Arena, "Agent Arena Leaderboard," snapshot dated 16 October 2026 (2,534,501 sessions, 52 models), accessed 19 October 2026. arena.ai/...agent
29. ^Arena Team, "Agent Arena: Causal Evaluation of Agents in the Real World," Arena blog, 4 June 2026, last updated 15 October 2026. arena.ai/...agent-arena-methodology
30. ^Xiaomi MiMo (@XiaomiMiMo), post on X, 16 October 2026. x.com/...2110944071875485925
31. ^1 ^2Arena, "Leaderboard Changelog," accessed 19 October 2026. arena.ai/...leaderboard-changelog
32. ^1 ^2 ^3 ^4 ^5 ^6 ^7Xiaomi MiMo Team, "Diagnosing and Mitigating Tool-Call Repetition in MiMo-V2.6," 27 September 2026. mimo.xiaomi.com/...mimo-v2-6-tool-call-repetition
33. ^1 ^2 ^3 ^4XiaomiMiMo, "MiMo-V2.6-Pro-MOPD" model card, Hugging Face, 27 September 2026. huggingface.co/...MiMo-V2.6-Pro-MOPD
34. ^1 ^2 ^3 ^4XiaomiMiMo, "MiMo-V2.6-Flash-MOPD" model card, Hugging Face, 27 September 2026. huggingface.co/...MiMo-V2.6-Flash-MOPD
35. ^1 ^2 ^3Oliver Chen and Anthony Ozerov, "Two-Thirds of MiMo v2.6's Coding Tasks Leak the Answer. Are Models That Exploit This Misbehaving?," Vals AI, 7 October 2026. vals.ai/...mimo-reward-hacking
36. ^Xiaomi MiMo API Platform, "API Pricing," updated 8 October 2026. mimo.mi.com/...pay-as-you-go
37. ^Xiaomi MiMo API Platform, "Model Deprecation," updated 9 October 2026. mimo.mi.com/...deprecate
38. ^Arena Team, "Arena Alignment Index," Arena blog, 8 October 2026. arena.ai/...ai-alignment-index
39. ^Arena (@arena), post on X on MiMo-V2.6-Pro's Code Arena: WebDev debut, 21 September 2026. x.com/...2102142912943489220
40. ^1 ^2Artificial Analysis, "Announcing Artificial Analysis Intelligence Index v5," 27 October 2026. artificialanalysis.ai/...sis-intelligence-index-v5
41. ^Artificial Analysis, "Artificial Analysis Progress Index v1 Methodology," accessed 29 October 2026. artificialanalysis.ai/...progress-index
42. ^Artificial Analysis, "LLM Leaderboard" (Intelligence Index v5.0 scores), accessed 29 October 2026. artificialanalysis.ai/...models

Show all 42 references