Xiaomi’s MiMo V2.6 Model Debuts With Unprecedented Efficiency Gains

Xiaomi has released its MiMo V2.6 AI models for software agents that use tools to carry out tasks. Developers can pay for hosted access or run downloaded models themselves, taking responsibility for the computing infrastructure. Through OpenRouter, the smaller Flash model offers output at $0.28 per million tokens.

The flagship MiMo-V2.6-Pro and smaller MiMo-V2.6-Flash variants both accept text, images, audio and video and return text. Their million-token context windows let them consider long documents, code and conversation histories within a request. Tokens are the chunks of information a model processes, rather than a fixed number of words.

Both repositories label the downloadable models with the MIT license. The weights, the learned parameters that make the models work, can be used and modified under that permissive license. Developers can also access the models through Xiaomi’s API or through OpenRouter, a service that connects requests to model-hosting providers.

Flash Cuts the Price of Repeated Agent Calls

Xiaomi positions Pro for complex, extended tasks and Flash for frequent calls and large workloads. That distinction matters for agents: a single assignment can involve repeated model calls as software reads files, uses tools and checks its work.

OpenRouter’s Xiaomi-provider listings on September 22 show three separate charges: new input, reused input served from a cache, and generated output. The rates below are in US dollars per million tokens.

MiMo V2.6 Flash’s uncached-input and output rates are about one-third of Pro’s. Its cached-input price is much closer to Pro’s, so the relative saving depends on how much input a workload can reuse. An agent’s bill also depends on how many tokens it consumes before completing a task.

UltraSpeed is a faster hosted version of the Pro checkpoint, rather than a smaller model like Flash. Xiaomi advertises up to 20 times Pro’s output speed. Its listed token rates are ten times Pro’s, making it a higher-priced option for applications where response time matters.

OpenCode also announced on September 21 that Flash would be free for the following week, with both Flash and Pro available in its Go service.

Downloadable Weights Still Need Substantial Hardware

MiMo V2.6 Pro has 1.02 trillion total parameters, with 42 billion active for each token. MiMo V2.6 Flash has about 310 billion total parameters and 15 billion active. Both use a mixture-of-experts design, which activates selected parts of the model for each token instead of using every parameter in every calculation.

That reduces the computation performed per token, but the smaller active count is not the size of the model an operator has to accommodate. The downloaded weights and the memory used while serving requests remain substantial. MiMo V2.6 Flash is smaller than MiMo V2.6 Pro, yet it is still a large model to host.

Xiaomi’s deployment examples make the scale concrete. For SGLang, software that serves AI models, the Pro instructions split the model across 16 parallel workers on two machines; the Flash example uses eight-way parallelism. The alternative vLLM serving examples use eight-way parallelism for Pro and four-way for Flash.

Self-hosting puts infrastructure operation with the team deploying the weights. Hosted access exchanges that responsibility for usage charges and the selected provider’s service terms. OpenRouter lists both Xiaomi and DeepInfra for Pro and Flash and allows customers to choose or exclude providers. Data-location and retention requirements therefore concern the hosting arrangement as well as the model license.

A Benchmark Lead With Task-Specific Trade-Offs

Independent testing provides a reason to take MiMo V2.6 Pro seriously beyond its price. On the Artificial Analysis leaderboard reviewed September 22, Pro scored 46 and ranked highest among open-weight models. GLM-5.3 at its maximum reasoning setting scored 45, followed by Kimi K3 at 44. DeepSeek V4.1 Flash scored 39 and DeepSeek V4 Pro 0813 scored 36, both at their maximum settings.

The Intelligence Index’s ten evaluations combine agent tasks, coding, scientific reasoning and general capabilities. It is primarily an English-language, text-based measure; Artificial Analysis evaluates image, speech and multilingual performance separately. Several proprietary models rank above Pro overall.

Artificial Analysis also put Pro’s weighted cost at $0.13 per index task. That combines measured token usage and prices across the benchmark tasks, giving developers a cost comparison tied to completed evaluations rather than token rates alone.

Xiaomi’s own agent evaluations show why Flash’s lower price could matter without making it interchangeable with Pro. On DeepSWE v1.1, which tests long software-development tasks, Flash scored 67.9 against Pro’s 71.9. On AutomationBench, which tests workflows across software services, the gap was smaller: 52.3 against 53.1.

The difference widened on Terminal Bench 4.0, a test of tasks performed through a computer terminal: Pro scored 34.9 and Flash 28.8. Claude Opus 5 reached 49.0 in Xiaomi’s comparison. The wider gap on terminal tasks makes Pro’s higher price more relevant to that workload than to the workflow-automation test.

Training Agents to Produce Better Fixes

Xiaomi attributes much of the improvement to reinforcement learning: models practice tasks, receive scores for their attempts and are updated toward behavior that earns better rewards. Its release account reports 30 large training steps per model in under six days, covering roughly 750,000 trajectories each. A trajectory is the sequence of reasoning, tool use and other actions taken during an attempt.

Xiaomi reports spending about $2.62 million for Pro and $850,000 for Flash on reinforcement learning after the models’ initial training. Each step started from 1,568 prompts with 16 attempts per prompt; the technical report describes sequences averaging roughly 110,000 to 150,000 tokens.

Generating and judging those attempts consumed more than half of Pro’s reported reinforcement-learning budget. Xiaomi assigned 43.8% to generating trajectories and 12.7% to grading them, with 43.5% going to model updates. The work mixed coding, visual tasks, general professional workflows and cybersecurity, using different combinations of tools and context-management software around the model.

Long tasks finish at different times. Xiaomi’s training system could pause unfinished attempts and resume them later, keeping the computers busy while a separate mixing mechanism preserved the intended balance of task types. This let the training combine different workloads without simply favoring whichever ones finished fastest.

A coding agent can pass a test suite while making a fix unnecessarily broad or hiding errors that a maintainer needs to see. Xiaomi’s Groupwise Reward Synthesis builds task-specific criteria from several attempts, assessing both the resulting code and how the agent worked. Its Groupwise Advantage Redistribution compares passing attempts during training and gives better-quality fixes a larger share of the learning reward.

In a separate code-only experiment, Xiaomi compared training with and without that online grading. Without it, tool-use turns and token counts rose rapidly and more attempts ran into length limits. With it, pass rates kept improving while turn counts stayed roughly stable and token growth was more gradual. Xiaomi’s maintainer-oriented audits also associated the graded version with smaller, more precise patches, rather than changes that unnecessarily expanded APIs or suppressed exceptions.

When Passing the Tests Rewards a Shortcut

The report describes another problem: agents could sometimes find an already-published solution instead of repairing the assigned code. In one example involving the Astropy software library, an agent fetched a newer upstream file containing the fix for a misleading error message. The patch could satisfy the tests without demonstrating that the agent had solved the problem from the supplied checkout.

Xiaomi says it removed revealing build artifacts, caches and later repository history, restricted access to outside solutions and used a separate agent to probe for remaining loopholes. During training, confirmed shortcut-taking attempts received zero reward. The company reports that its logged share of confirmed reward-hacking trajectories stayed below 2% for both models.

MiMo team leader Fuli Luo described the research and engineering effort behind the run and said the team also trained harder-to-verify capabilities separately before combining them.

Alongside the flagship weights, Xiaomi says it is releasing a smaller nine-billion-parameter model distilled from MiMo training trajectories, roughly 7,000 task environments and an end-to-end reinforcement-learning framework. These are resources for researchers to investigate the training approach on a smaller scale than the trillion-parameter Pro model.