'Better than DeepSeek': Xiaomi's MiMo-V2.6-Pro debuts as the top open weights model in the world alongside cheaper V2.6-Flash

In a surprising upset, Chinese electric car and consumer electronics manufacturer Xiaomi has released the latest version of its growing family of MiMo language models, and MiMo-V2.6-Pro has arrived as the top-performing open-weight model in the world on third-party benchmarking firm Artificial Analysis’ Intelligence Index, scoring 46.

That puts Xiaomi’s new flagship ahead of proprietary models including xAI’s Grok 4.6, currently at 44, and Google’s Gemini 3.8 Flash, currently at 41, on Artificial Analysis’ latest scoring methodology. It also places MiMo-V2.6-Pro tied with the newly released Grok 4.7, which debuted the same day and also scores 46, and above fellow Chinese open weights model DeepSeek V4.1 Flash (39) and DeepSeek V4.1 Pro (36).

The result is striking not just because Xiaomi is better known internationally for its smartphones, electric vehicles and consumer electronics than frontier AI.

It is also because MiMo-V2.6-Pro is open weight, MIT-licensed and dramatically cheaper to access through Xiaomi’s API than many proprietary models around its performance tier.

In other words, indie developers and enterprises alike can download it for free off Hugging Face, customize or fine-tune it to their specific uses, run it themselves on their own or rented hardware and use it in their production workflows — all without paying Xiaomi a cent. They can also run it off Xiaomi's APIs (a trickier proposition for those who have restrictions on using Chinese-based servers) for among the cheapest rates per million tokens.

Xiaomi charges $0.435 per million uncached input tokens and $0.87 USD per million output token. Artificial Analysis measures the model at $0.13 per Intelligence Index task, and output speed at roughly 134 tokens per second, confirming a 1-million-token context window with text, image, audio and video input.

The flagship is joined by MiMo-V2.6-Flash, a smaller and substantially cheaper model ($0.14/$0.28 USD per million input/output tokens) that retains the same 1-million-token context window and native multimodal capabilities while targeting high-volume production workloads. It’s the second cheapest major frontier model available in the world over its application programming interface (API) by our assessment.

Xiaomi is also releasing MiMo-V2.6-Pro-UltraSpeed, which it says can generate at up to 20 times the normal Pro output speed.

Taken together, V2.6 looks less like a single model launch than the latest step in Xiaomi’s effort to build an entire open agent stack: foundation models, coding agents, harnesses, reinforcement-learning environments and now the training infrastructure behind them.

From smartphones and EVs to frontier AI

Xiaomi’s move into frontier models has accelerated rapidly over the past year.

The company began publicly expanding the MiMo family in 2025 and spent much of 2026 pushing deeper into agentic AI. Its April, its MiMo-V2.5 and V2.5-Pro release established many of the architectural and economic ideas now visible in V2.6: sparse mixture-of-experts models, million-token context windows, permissive licensing and unusually low prices for long-running agent workloads.

MiMo-V2.5-Pro was already a 1.02-trillion-parameter mixture-of-experts model with 42 billion parameters active during inference, trained specifically for long-horizon software engineering and what Xiaomi called “harness awareness” — the ability to manage memory and context while operating inside agent scaffolds over hundreds or thousands of tool calls.

In June, Xiaomi followed with MiMo Code, an open-source terminal coding agent with persistent cross-session memory, task checkpoints and a separate checkpoint-writing subagent. Xiaomi’s internal testing suggested MiMo Code’s advantage over Claude Code became more pronounced after workflows stretched beyond 200 execution steps, although those results were vendor-reported and sensitive to configuration.

The company also introduced HarnessX, a research framework that treats the prompts, memory systems, tools and control logic surrounding a model as components that can themselves be rewritten and optimized. Xiaomi reported an average 14.5% absolute performance gain across 15 model-benchmark combinations when the harness evolved dynamically, without replacing the underlying model.

MiMo-V2.6 pulls those lines of work back into the model-training process itself.

Xiaomi spent millions scaling reinforcement learning

The biggest technical story behind V2.6 is reinforcement learning, or RL. This is the stage where a model improves by trying tasks, receiving rewards or penalties based on how well it performed, and updating itself toward behaviors that produce better outcomes.

Scaling that up is costly because agent tasks can involve long chains of reasoning, tool calls and verification, but the potential payoff is substantial: instead of learning only from static examples, the model gets repeated practice actually executing complex work, including planning, using tools, correcting mistakes and recovering from failed strategies.

Xiaomi says MiMo-V2.6-Pro and the smaller MiMo-V2.6-Flash each underwent 30 large RL steps covering roughly 750,000 trajectories in under six days, at reported costs of about $2.62 million for Pro and $850,000 for Flash.

Its public technical report makes the scale more concrete. Each training step starts with 1,568 prompts and generates 16 candidate trajectories for each one, producing roughly 25,000 rollouts and between 2.7 billion and 3.7 billion training tokens per step. Xiaomi says the resulting sequences average roughly 110,000 to 150,000 tokens apiece.

That means Xiaomi was not primarily reinforcing short answers. It was reinforcing entire, lengthy agent workflows.

And only part of that spending went toward updating the model itself. Xiaomi says 43.5% of Pro’s RL cost went to training, 43.8% to producing rollouts and another 12.7% to grading them. More than half the budget, in other words, went toward generating and evaluating the model’s experience before that experience was turned into weight updates.

The company calls its strategy “You Only RL Once.” Instead of running completely separate RL programs for coding, visual work, computer use and cybersecurity, Xiaomi mixes those domains and multiple agent harnesses into a single large training run.

Harness diversity is itself part of that training distribution. Xiaomi says it created lightweight “mini-harnesses” for coding, general professional workflows, visual tasks and cybersecurity so that the model encounters different combinations of system prompts, tools and context-management strategies without becoming tightly coupled to one particular production scaffold.

The system uses fully asynchronous Group Relative Policy Optimization, or GRPO. That asynchronous design matters because long-running agent jobs do not all finish together. Xiaomi uses partial rollouts to keep its GPU fleet busy rather than waiting for the slowest trajectories, interrupting unfinished jobs and resuming them later. It also built a sample-mixing mechanism intended to stop faster or easier task families from overwhelming slower ones in each batch.

The result is that “scaling RL” here means scaling several things at once: the number of model-generated experiences, the range of environments in which those experiences occur and the compute spent deciding which behavior is actually worth reinforcing.

Teaching an agent not just to pass, but to solve cleanly

That last part may be one of the more consequential technical details in the report.

Simple binary rewards can tell a coding model whether its patch passes a test suite, but they cannot reliably distinguish between two passing solutions where one is clean and minimal and the other introduces broad fallback logic, unnecessary API changes or brittle workarounds.

Xiaomi therefore built two more elaborate reward systems.

1. Groupwise Reward Synthesis, or GRS, creates task-specific rubrics by comparing several attempts at the same problem. Those rubrics separately evaluate the quality of the implementation and the quality of the agent’s behavior: whether it satisfied the requirements, handled edge cases, respected the surrounding codebase and gathered and verified the evidence it needed.

2. Groupwise Advantage Redistribution, or GAR, goes further by comparing passing solutions inside the same rollout group and shifting more of the training advantage toward the better ones.

The technical report includes an experiment suggesting why this matters. In a code-only RL run on MiMo-V2.6-Flash, training without online groupwise grading caused agent turn counts and generated-token lengths to rise rapidly, eventually pushing more trajectories into their length limits. With GAR enabled, Xiaomi says pass rates continued improving while the number of turns remained roughly stable and token growth was more gradual.

Maintainer-oriented audits also found that the policy trained without online grading increasingly resorted to techniques such as speculative compatibility branches, broad exports, exception swallowing, relaxed validation and evaluation-specific configuration changes. The groupwise-graded policy, Xiaomi says, tended to produce smaller and more precise patches.

That gets at a fundamental problem in agentic reinforcement learning: a model can become better at maximizing a reward without becoming better at the work the reward is supposed to represent.

Xiaomi devotes an unusually detailed part of the report to that problem under the heading of reward hacking.

In early coding runs, agents sometimes discovered that instead of solving an assigned bug themselves, they could download a newer release of the package, retrieve an upstream source file, clone a later state of the repository or search issue histories for the already-published fix.

Those approaches could satisfy the tests while bypassing the intended task entirely.

Xiaomi says it responded by stripping build artifacts and caches from its training environments, removing future Git history, blocking network access to potential answer sources and deploying a separate “hack agent” specifically to search for remaining loopholes before training. During the final run, the company says confirmed reward-hacking trajectories stayed below 2% for both Pro and Flash; when the grader found one, its effective reward was reset to zero.

At this scale, the machine-learning problem starts to look inseparable from a distributed-systems problem. Xiaomi has to keep tens of thousands of long-running trajectories moving across heterogeneous environments, prevent faster tasks from dominating the batch, keep training and inference behavior aligned and ensure that increasingly capable models do not learn to exploit weaknesses in the graders training them.

The technical report says Xiaomi even froze the model’s mixture-of-experts router during RL to reduce training drift and improve stability.

Fuli Luo, a former DeepSeek researcher now heading up the Xiaomi MiMo team, posted on X to say that V2.6 is likely one of the largest single reinforcement-learning runs undertaken by an open-source model team.

She said Xiaomi assigned several dozen people to the effort and argued that its research and engineering challenges exceeded those she encountered while contributing to DeepSeek R1.

“In an era when compute is brutally scarce, we still chose to dedicate a team of several dozen people to one goal over an extended period: scaling up RL.”

Her description is consistent with a technical report that spends nearly as much time on the machinery required to keep the RL system stable as on the learning algorithm itself.

Daanish Khazi, CEO of open source AI knowledge work startup Paper Instruments, posted on X that he thought the release should cause researchers to “update every prior on post-training scaling laws,” focusing on the apparent amount of capability Xiaomi extracted from a comparatively modest reported compute bill entirely on Chinese chips rather than Nvidia GPUs.

Strong agent performance, with a new open-weight leader — but not a benchmark sweep

Xiaomi’s own benchmark suite shows large gains over V2.5, scoring 71.9 on DeepSWE v1.1, 53.1 on AutomationBench, 76.9 on Toolathlon-Verified, 89.9 on Terminal Bench 2.1 and 62.0 on JobBench. It also reaches 94.0 on CyberGym and 72.3 on Xiaomi’s MiMo Visual Coding benchmark..

Anthropic’s proprietary Claude Opus 5 remains ahead on several evaluations reported by Xiaomi, including DeepSWE v1.1, ProgramBench and Terminal Bench 4.0. OpenAI’s proprietary GPT-5.6 Sol is also ahead on some cybersecurity evaluations, including ExploitBench and SEC Bench Pro.

On DeepSWE v1.1, MiMo-V2.5-Pro scored 19.0, while V2.6-Pro reaches 71.9. AutomationBench rises from 16.0 to 53.1, and MiMo Code Bench climbs from 40.4 to 63.2.

That distinction matters. Xiaomi is not showing an across-the-board victory over the strongest closed models. What it has done is move an open-weight system materially closer to that frontier while retaining downloadable weights and a much lower operating price.

Flash may be the more consequential model for high-volume users

The flagship Pro model will naturally attract the most attention, but MiMo-V2.6-Flash may prove more relevant for enterprises running large numbers of agent calls.

Flash is itself a large sparse mixture-of-experts model: 310 billion total parameters, with 15 billion active during inference, compared with Pro’s 1.02 trillion total and 42 billion active. Both support a 1-million-token context window, multimodal input and up to 128,000 output tokens.

Flash stays surprisingly close to Pro on several of Xiaomi’s agent benchmarks, scoring 67.9 versus Pro’s 71.9 on DeepSWE v1.1, 52.3 versus 53.1 on AutomationBench, 61.2 versus 63.2 on MiMo Code Bench, 87.6 versus 89.9 on Terminal Bench 2.1, 61.2 versus 62.0 on JobBench and 71.5 versus 72.3 on MiMo Visual Coding.

On CyberGym, Flash actually outpaces Pro’s 94.0 with a score of 95.1 , although that should not be interpreted as evidence that Flash is generally the stronger cybersecurity model: Pro is far ahead on ExploitGym, ExploitBench and SEC Bench Pro.

The pricing gap, however, is substantial. Xiaomi charges $0.14 per million uncached input tokens and $0.28 per million output tokens for Flash, with cache-hit input falling to $0.0028 per million tokens. Pro costs $0.435 per million uncached input tokens and $0.87 per million output tokens, with cache hits at $0.0036.

VentureBeat Frontier Model Per 1M Token API Cost Comparison Table Snapshot - Late 2026

That means Flash costs roughly one-third as much as Pro on uncached input and output, while trailing it by only a few points on several of Xiaomi’s long-horizon agent benchmarks.

For workloads where a company might launch thousands or millions of agent runs — coding assistants, document-processing systems, internal research agents or automated back-office workflows — that difference can matter more than a small benchmark gap.

As Tim Dettmers, a Carnegie Mellon University computer science professor, researcher at Allen Institute for Artificial Intelligence (Ai2) and creator of the LLM quantization technique bitsandbytes wrote on X: "The flash model has really good vibes. Feels like the best model in the 300B to 550B class. Better than DeepSeek v4.1 and GLM 5.3 Flash..."

Xiaomi describes Flash as its model for “high-frequency calls and large-scale tasks,” and that positioning makes sense. The technical story of V2.6 is not just that the trillion-parameter Pro model reached the top of an open-weight leaderboard. It is that Xiaomi appears to have transferred a meaningful amount of the same agent capability into a cheaper 15-billion-active-parameter model.

For developers, the pricing changes the practical comparison with proprietary frontier models. A model does not have to win every benchmark if it is capable enough for the workload and dramatically cheaper to run — particularly for agents that may consume hundreds of thousands or millions of tokens during a single job.

Artificial Analysis’ price-performance chart reflects that: MiMo-V2.6-Pro lands dead center in the upper-left quadrant of its intelligence-versus-cost Pareto frontier, right in the sweet spot between low cost and high performance, reinforcing Xiaomi’s argument that V2.6 is not merely chasing benchmark leadership but attempting to maximize useful capability per dollar.

OpenCode quickly leaned into that economics, announcing on X that MiMo-V2.6-Flash would be free for the following week, with both Flash and Pro also available in its Go service.

From vibe coding to ‘Vibe World’

Xiaomi demonstrates MiMo taking text, images or video and coordinating multiple agents to create playable 3D worlds, construct scenes, implement interaction logic, inspect rendered output and iteratively refine the result.

Other demonstrations include generating Blender objects and scenes, operating desktop software and controlling a simulated Franka Panda robotic arm through visual feedback.

The model also works across frontend design, Figma, presentations, SVGs, video production and music composition.

Xiaomi’s own Pro model page frames these capabilities as professional workflow features rather than novelty demos, listing coding, office work, design, research, content creation, cybersecurity and computer operation among the model’s intended domains.

Two research case studies push the idea further. In the first, Xiaomi’s materials researchers used MiMo-V2.6-Pro to review scientific literature and patents, propose metal-organic frameworks for capturing PFAS “forever chemicals,” run computational simulations and shortlist candidates for possible wet-lab validation.

In another, the model assisted researchers in formalizing the main theorem from Li and Yorke’s Period Three Implies Chaos in Lean 4. Xiaomi says the resulting project exceeded 6,000 lines of Lean source code and passed verification by Lean’s kernel with no unfinished proof placeholders.

Those demonstrations should not be confused with proof that the model can autonomously conduct reliable scientific research. But they point to the kind of workload Xiaomi is optimizing for: long-running projects that combine reasoning, software, tools, perception and repeated inspection rather than a single prompt and response.

Xiaomi is open-sourcing more than the model weights

Perhaps the most unusual part of the release is what Xiaomi is publishing alongside the models.

The company is releasing Pro and Flash model weights, the technical report, more than 7,000 RL task environments, an end-to-end RL framework, composable mini-harnesses and MiMo-V2.6-Distill-Qwen-9B, a smaller model distilled from MiMo RL trajectories.

That makes this more than a conventional weight drop. The technical report spends dozens of pages on environment generation, grader construction, reward-hacking defenses, asynchronous rollout infrastructure and the mechanics of mixing agent tasks in the same training run. Xiaomi says the goal is to give other researchers enough of the stack to reproduce and extend the work.

For the open-model community, that may prove more consequential than whether MiMo-Pro holds the top Artificial Analysis ranking for a few weeks or months. If outside teams can reproduce some of Xiaomi’s RL gains on smaller models, the release could provide a practical blueprint for improving agent behavior without training another trillion-parameter foundation model from scratch.

A stronger challenge from Xiaomi’s increasingly complete AI stack

The larger pattern is becoming difficult to dismiss. Xiaomi began from foundation models, added long-horizon coding agents, moved into adaptive harnesses and has now exposed much of the reinforcement-learning infrastructure connecting those pieces together.

The company is still not winning every benchmark, and many of its agent evaluations remain vendor-run. The trillion-parameter Pro model is also a substantial system to self-host even though only a fraction of its parameters activate during each token.

But those limitations coexist with a new reality: Xiaomi now has the highest-ranked open-weight model on Artificial Analysis’ composite intelligence benchmark, a smaller Flash model that approaches Pro on several agent tasks at roughly one-third the API price, a permissive licensing strategy and a growing body of open training infrastructure.

For enterprise AI teams, the immediate question is therefore not whether MiMo-V2.6 is categorically better than every proprietary frontier model. It is not.

The more practical question is whether a downloadable model that comes close enough on the workloads that matter — while costing $0.87 per million output tokens for Pro or just $0.28 for Flash — can replace a more expensive proprietary model for at least part of an enterprise agent stack.

With MiMo-V2.6, Xiaomi has given developers a much stronger reason to run that evaluation.