On September 22, 2026, Xiaomi released Pro and Flash versions of its large-scale AI model, "MiMo-V2.6." The flagship MiMo-V2.6-Pro scored 46 on the third-party evaluator Artificial Analysis's "Intelligence Index," placing it at the top among open-weight models as of the September 23 check. Xiaomi's own official announcement puts the figure at 46.32, including decimals.
The model weights are released under the MIT license, and the model can take text, images, video, and audio as input, handling a context window of up to 1 million tokens.
However, reading this announcement simply as "Xiaomi topped an AI ranking" misses the point.
MiMo-V2.6-Pro did not surpass every AI model out there. What deserves more attention is that Xiaomi opened up to the outside world the large-scale reinforcement learning (RL) infrastructure it spent nearly half a year building — not just the model weights, but the training environments and code as well.
And when it comes to actual use, what matters even more than Pro — which exceeds 1 trillion parameters — is Flash, which cuts API pricing to roughly a third without a major drop in performance.
MiMo-V2.6-Pro Leads Among Open-Weight Models, Not the Entire AI Field
The Intelligence Index v4.3.2 currently used by Artificial Analysis is a composite metric combining ten types of evaluations, including knowledge work, agentic task automation, device operation, scientific coding, and long-context reasoning.
As of the September 23 check, MiMo-V2.6-Pro scored 46 — first place among open-weight models.
Meanwhile, if closed models are included, Claude Opus 5's maximum reasoning setting scored 51, and GPT-5.6 Sol's maximum setting scored 47. Alibaba's Qwen3.8 Max scored 45, the open-weight Kimi K3 scored 44, and DeepSeek V4.1 Flash scored 39.
データを表で見る
| Intelligence Index (points) | |
|---|---|
| Claude Opus 5 max | 51 |
| GPT-5.6 Sol max | 47 |
| MiMo-V2.6-Pro | 46 |
| Qwen3.8 Max | 45 |
| Kimi K3 max | 44 |
| DeepSeek V4.1 Flash | 39 |
Once price is factored in, MiMo-V2.6-Pro's position becomes even more distinctive.
| Model | Weights | Intelligence Index | AA Cost per Task | Context |
|---|---|---|---|---|
| Claude Opus 5 (max) | Closed | 51 | $5.86 | 1M |
| GPT-5.6 Sol (max) | Closed | 47 | $1.99 | 1M |
| MiMo-V2.6-Pro | Open, MIT | 46 | $0.13 | 1M |
| Qwen3.8 Max | Closed | 45 | $5.41 | 984K |
| Kimi K3 (max) | Open | 44 | $2.00 | 1M |
| DeepSeek V4.1 Flash (max) | Open | 39 | $0.27 | 1M |
The "cost per task" figure here is different from the API price per million tokens.
It is a weighted average of the cost required to complete one task — including input, cache, reasoning, and output tokens — when Artificial Analysis actually ran the evaluation.
MiMo-V2.6-Pro comes in at $0.13, considerably lower than models with a similar Intelligence Index. In other words, what stands out is not just the score of 46, but how cheaply that 46 is achieved.
That said, this is not an absolute performance metric either. The numbers shift if Artificial Analysis's evaluation set or a model's reasoning settings change.
Calling MiMo-V2.6-Pro "the strongest AI" would not be accurate; it is more fair to describe it as "the top-scoring open-weight model on Artificial Analysis v4.3.2 as of September 23, 2026."
Even by Xiaomi's Own Benchmarks, It Doesn't Beat Closed Models Across the Board
This nuance becomes clearer when looking at the benchmarks Xiaomi itself has published.
The company's model card compares MiMo-V2.6 against Claude Opus 5, GPT-5.6 Sol, and Claude Fable 5 on the same agentic evaluations.
| Benchmark | MiMo-V2.6 Pro | MiMo-V2.6 Flash | Claude Opus 5 | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|---|---|---|
| DeepSWE v1.1 | 71.9 | 67.9 | 74.0 | 73.0 | 70.0 |
| AutomationBench v1.0.6 | 53.1 | 52.3 | 50.3 | 45.8 | 46.2 |
| Terminal Bench 4.0 | 34.9 | 28.8 | 49.0 | 39.9 | 42.4 |
| Terminal Bench 2.1 | 89.9 | 87.6 | 89.1 | 88.8 | 84.3 |
| JobBench | 62.0 | 61.2 | 65.7 | 45.4 | 57.4 |
| ExploitBench | 47.9 | 25.3 | 70.0 | 78.5 | 78.0 |
| MiMo Visual Coding | 72.3 | 71.5 | 70.0 | 73.4 | 69.1 |
On DeepSWE v1.1, which measures extended software development work, Pro scored 71.9 versus 74.0 for Claude Opus 5 and 73.0 for GPT-5.6 Sol. MiMo also trails closed models on Terminal Bench 4.0.
On the other hand, on AutomationBench, which covers business task operation, Pro scored 53.1 and Flash 52.3, both surpassing the comparison models, and on Terminal Bench 2.1 Pro reached 89.9.
In other words, MiMo-V2.6 is better understood not as an all-around model that improves uniformly across every task, but as a model whose capabilities as an "agent" — one that uses tools to carry out extended work — have advanced rapidly.
25,088 Trajectories Per Update: What Xiaomi Really Scaled Was "Experience"
MiMo-V2.6-Pro is a Mixture of Experts (MoE) model with 1.02 trillion total parameters and 42 billion active parameters per token.
Flash has 309 billion total parameters and 15 billion active parameters. Both models have a maximum context of 1 million tokens and accept text, image, video, and audio input.
| MiMo-V2.6 Pro | MiMo-V2.6 Flash | |
|---|---|---|
| Total parameters | 1.02T | 309B |
| Active parameters | 42B | 15B |
| Max context | 1M tokens | 1M tokens |
| Pretraining volume | 30T tokens | 48T tokens |
| Architecture | MoE | MoE |
| Multimodal | Yes | Yes |
The pretraining volume is also substantial.
According to the technical report, Flash was trained on 48 trillion tokens and Pro on 30 trillion tokens. For Flash, that breaks down into 26 trillion in the text stage and 22 trillion in the multimodal stage; for Pro, it was 27 trillion and 3 trillion respectively.
During pretraining, the context length was extended from 32K to 256K, and then further stretched to 1 million tokens during a subsequent mid-training phase that incorporated agentic work trajectories.
But what really sets MiMo-V2.6 apart comes after that.
For each RL step, Xiaomi selected 1,568 tasks and generated 16 rollouts from each.
Doing the simple math, that comes to 25,088 agent trajectories per update.
Since Pro and Flash each went through 30 steps, roughly 750,000 trajectories were generated per model. The number of training tokens processed per step reached into the billions — around 2.7 to 3.7 billion according to the technical report, and 3.5 to 3.7 billion tokens by the later stages according to the official announcement.
This is quite different from the typical RL approach of "give a question, score the answer."
A single trajectory involves extended work — reading a codebase, searching for a file, invoking tools, rewriting code, running tests, observing a failure, and fixing it again.
The average trajectory reportedly reached 110,000 to 150,000 tokens.
What Xiaomi scaled up was not simply parameter count, but the sheer volume of "trial and error" experience the model was put through.
Roughly $2.62 Million for RL Alone on Pro — and Over Half of That Wasn't Even "Training"
This RL run finished in under six days.
According to the official announcement, the cost for Pro was about $2.62 million, and for Flash about $850,000. Over the 30 steps, DeepSWE v1.1 rose from 58.4 to 72.57 for Pro, and from 48.8 to 65.68 for Flash.
The breakdown of costs is even more interesting.
For Pro, 43.8% went to rollout generation, 43.5% to actual parameter training, and 12.7% to the grader that scores answers.
データを表で見る
| Cost share (%) | |
|---|---|
| Rollout generation | 43.8 |
| Model training | 43.5 |
| Scoring/evaluation | 12.7 |
In other words, more than half of the RL cost was spent not on "the computation that updates the model's weights," but on generating the experience used for training and judging the quality of that experience.
Herein lies an important design philosophy behind MiMo-V2.6.
In conventional RL, it's easy to use a simple pass/fail binary reward based on whether a test succeeded. But even among answers that pass the same test, there's a quality difference between a solution that made an unnecessarily large number of changes and one that fixed the problem correctly with minimal changes.
MiMo-V2.6 introduces Groupwise Reward Synthesis (GRS) and Groupwise Advantage Redistribution (GAR), which compare multiple trajectories generated from the same task and evaluate quality even among successful answers.
By investing compute resources into the evaluation side as well, the system doesn't just increase "successful actions" — it steers rewards toward shorter solutions with fewer unnecessary changes.
This is also an example of the competitive focus shifting from simply making models bigger toward the question of "how much high-quality experience can be given to the model."
Mixing Code, Business Task Operation, Vision, and Cybersecurity Into a Single RL Run
Xiaomi calls this training method "You Only RL Once."
Rather than finishing separate models for code generation, PC operation, visual agents, and cybersecurity, the company mixes multiple types of tasks and multiple agent execution frameworks — so-called harnesses — into a single RL run.
To prevent the model from memorizing only specific system prompts or tool configurations, different harnesses were treated as part of the training distribution.
However, at this scale, simple parallel processing doesn't work.
If tasks that take a few minutes and tasks that take dozens of minutes are placed in the same batch, only the short tasks finish first.
So Xiaomi introduced partial rollouts, which resume unfinished trajectories later, and a sample mixer, which maintains the ratio of different task types.
Because MoE architectures can suffer from uneven training load concentrating on specific experts, the router that selects experts was fixed during RL.
Countermeasures against reward hacking were also needed.
If the model learns to exploit weaknesses in the grader instead of actually solving the task correctly, then large amounts of RL compute end up working against the goal.
Xiaomi combines reward design, adversarial evaluation, anomaly detection, and cross-checking with multiple verifiers.
In large-scale agentic RL, the bottleneck is often not the model itself but the system that runs task supply, scheduling, tool execution, graders, and inference infrastructure together.
MiMo-V2.6 can be described as a model that squarely tackled that problem.
Flash May Have More Market Impact Than Pro
And from a practical standpoint, MiMo-V2.6-Flash is the more interesting release.
While Pro has 1.02 trillion total parameters with 42 billion active, Flash is considerably smaller, at 309 billion total with 15 billion active.
Even so, on Xiaomi's key agentic evaluations, Flash keeps the gap with Pro down to just a few points on many items.
- MiMo-V2.6 Pro
- MiMo-V2.6 Flash
データを表で見る
| MiMo-V2.6 Pro (score) | MiMo-V2.6 Flash (score) | |
|---|---|---|
| DeepSWE v1.1 | 71.9 | 67.9 |
| AutomationBench | 53.1 | 52.3 |
| Terminal Bench 2.1 | 89.9 | 87.6 |
| JobBench | 62 | 61.2 |
| MiMo Visual Coding | 72.3 | 71.5 |
| Terminal Bench 4.0 | 34.9 | 28.8 |
| ExploitBench | 47.9 | 25.3 |
Looking at the gap in table form makes the pattern clearer.
| Metric | Pro | Flash | Gap |
|---|---|---|---|
| DeepSWE v1.1 | 71.9 | 67.9 | 4.0 |
| AutomationBench | 53.1 | 52.3 | 0.8 |
| Terminal Bench 2.1 | 89.9 | 87.6 | 2.3 |
| JobBench | 62.0 | 61.2 | 0.8 |
| MiMo Visual Coding | 72.3 | 71.5 | 0.8 |
| Terminal Bench 4.0 | 34.9 | 28.8 | 6.1 |
| ExploitBench | 47.9 | 25.3 | 22.6 |
However, the gap in API pricing is far wider than the gap in performance.
| Model | Uncached input / 1M tokens | Output / 1M tokens | Cache-hit input |
|---|---|---|---|
| MiMo-V2.6-Pro | $0.435 | $0.87 | $0.0036 |
| MiMo-V2.6-Flash | $0.14 | $0.28 | $0.0028 |
| Pro UltraSpeed | $4.35 | $8.70 | $0.036 |
- MiMo-V2.6 Pro
- MiMo-V2.6 Flash
データを表で見る
| MiMo-V2.6 Pro (USD) | MiMo-V2.6 Flash (USD) | |
|---|---|---|
| Uncached input | 0.44 | 0.14 |
| Output | 0.87 | 0.28 |
Flash's standard input price is 32.2% of Pro's — a discount of roughly 68%. Output pricing follows almost the same ratio.
This is not a difference that can be ignored in agentic use cases.
For a single response, a difference of a few dozen cents is negligible.
But in setups where an agent plans, calls multiple sub-agents, reads code, runs tools, and retries after failures, a single user action can trigger dozens or even hundreds of model calls.
In that case, being able to try a model that maintains a sufficient success rate roughly three times as often can improve the overall system's success rate more than squeezing out a few extra points of peak performance.
That said, it would be risky to think of Flash as simply "a cheaper Pro."
On Terminal Bench 4.0 the gap widens to 34.9 vs. 28.8, and on ExploitBench to 47.9 vs. 25.3.
On certain tasks, the difference in model scale shows up clearly.
In real-world deployment, one could imagine a setup where Flash handles routine processing and large-scale parallel attempts, while only the highest-difficulty or highest-failure-cost work is escalated to Pro.
But that decision should be made not based on public benchmarks, but on task-by-task success rates measured with one's own tools, prompts, data, and harness.
What It Means to Release Over 7,000 RL Environments
Another unusual aspect of this release is that it didn't stop at the model weights.
Alongside Pro and Flash, Xiaomi released a technical report, an end-to-end RL framework, a lightweight mini-harness, and more than 7,000 RL task environments.
These environments cover software development, vulnerability reproduction, knowledge-intensive work, and web design/development, among others.
The RL framework is built on top of verl, uni-agent, and mini-swe-agent, handling everything from environment operation and trajectory collection to reward evaluation and policy optimization.
Furthermore, for researchers who cannot retrain the massive Pro or Flash models, Xiaomi also prepared "MiMo-V2.6-Distill-Qwen-9B."
This is a model created by supervised fine-tuning of Qwen3.5-9B on data generated by MiMo. The SFT data totaled 77.4 billion tokens, of which 27.2 billion were actual loss-bearing tokens used in the loss calculation.
Xiaomi reports that starting from this 9B model and training it further on the published RL environments, it outperformed the SFT-only version on every reported metric.
| Benchmark | SFT | After RL | Improvement |
|---|---|---|---|
| SWE-bench Verified | 61.1 | 66.2 | +5.1 |
| MiMo Cyber Bench | 31.3 | 47.0 | +15.7 |
| Terminal Bench 2.1 | 37.1 | 52.8 | +15.7 |
| MiMo Visual Coding | 64.0 | 72.4 | +8.4 |
- SFT only
- After RL
データを表で見る
| SFT only (score) | After RL (score) | |
|---|---|---|
| SWE-bench Verified | 61.1 | 66.2 |
| MiMo Cyber Bench | 31.3 | 47 |
| Terminal Bench 2.1 | 37.1 | 52.8 |
| MiMo Visual Coding | 64 | 72.4 |
This is where the long-term value of this release lies.
Even if a 1.02-trillion-parameter model is released publicly, researchers who cannot secure thousands of GPUs cannot reproduce the same RL run.
But when a 9B model, task environments, graders, harness, and training code are all provided together, questions like "does mixing multiple harnesses in training actually improve generalization to unseen environments?" or "does groupwise grading help with other models too?" can be tested and disentangled at a much smaller compute scale.
Artificial Analysis's top spot will be lost once the next model comes along.
But if this training environment and RL methodology gets reused by other developers, MiMo-V2.6's influence could outlast its ranking.
"42 Billion Active Parameters" Doesn't Mean It Behaves Like a 42B-Parameter Model
Being open-weight doesn't mean it's easy to operate, either.
Pro activates only 42 billion parameters per token, but the model as a whole contains 1.02 trillion parameters.
While the number of active parameters matters for thinking about inference compute, the full set of weights must still be handled when storing the model and distributing experts across GPUs.
The SGLang configuration for Pro described in Xiaomi's official model card uses 16-way tensor parallelism and expert parallelism, and assumes execution across multiple GPUs and multiple nodes.
Don't look at the "42 billion active" figure alone and assume it can be loaded onto a single GPU the way an ordinary 42B-parameter model could.
Because of this, for many users "open-weight" doesn't automatically translate into "cheaper to run on your own servers."
You need to compare API pricing, GPU procurement costs, utilization rates, cache hit rates, latency, data residency, and operational staffing all together.
Compare by "Cost to Complete One Task Successfully," Not Price per Token
Another shift that MiMo-V2.6 highlights is how AI model pricing itself should be compared.
Looking only at whether it's $0.14 or $0.435 per million tokens doesn't tell you much about the economics of running agents.
What matters is how many tokens it takes, how many times it fails, how many retries occur, and how long it takes to complete a single job.
For instance, even a cheap model can end up costing more than an expensive model that finishes in one attempt, if it needs three retries to succeed.
On the other hand, for tasks like those where Flash's gap with Pro is small, the difference in unit price translates directly into a large difference in overall cost.
This is exactly the distinction Artificial Analysis is trying to capture when it calculates, separately from the per-million-token price, a figure of "$0.13 per task on the Intelligence Index" for MiMo-V2.6-Pro.
According to the same company's measurements, MiMo-V2.6-Pro's output speed also reaches roughly 130 tokens per second.
Going forward, when choosing an agentic model, the question
"which model is the smartest?"
may matter less than the question
"how much did it cost per successfully completed task?"
The Real Test for MiMo-V2.6 Is Just Beginning
MiMo-V2.6-Pro's Artificial Analysis Intelligence Index score of "46" is an easy headline.
But that number alone won't last long. AI model rankings get rewritten the moment the next strong model appears.
What matters more here is that Xiaomi generated roughly 750,000 extended agent trajectories per model, poured significant compute not just into rollouts but into the grader as well, mixed different types of tasks and harnesses into a single RL run, and released the supporting environments and code that made it all possible.
And on the commercial side, Flash maintains performance close to Pro on many evaluations while cutting standard API pricing to roughly a third.
Two things are worth watching going forward.
One is whether Flash can maintain a "task success rate" close to Pro's in actual development, research, and business automation work.
The other is whether researchers outside Xiaomi can reproduce similar improvements using the published 9B model and the more than 7,000 RL environments.
Only once those questions are answered can we say that MiMo-V2.6's six-day large-scale RL run was not just "another high-performance giant model being built," but rather an event in which the very process of how to train agents was handed over to the outside world.
That could end up mattering far longer than any ranking.