MiMo-V2.6-Pro is the latest release from Xiaomi, and it is trying to do something unusual: sell an AI model that is both reasonably priced and reasonably smart, even if it talks a lot and thinks slowly. The model, which was released on September 21, 2026, is a reasoning model with a 1M tokens context window. It supports text, image, speech, and video input and generates text output. Its Intelligence Index score is 46, placing it well above average among comparable models with a median of 18.
The catch is the cost. MiMo-V2.6-Pro runs $0.43 per 1M input tokens and $0.87 per 1M output tokens. It also takes over 2 seconds to produce its first token and generates a hefty 140M output tokens whenever it is evaluated. That is a lot of talking for a model that is supposed to be priced reasonably.
Intelligence at a Glance
The Intelligence Index is the number Xiaomi built the model’s reputation around. A score of 46 puts MiMo-V2.6-Pro above the median of 18 among comparable models. The index tests models across reasoning, knowledge, mathematics, and coding, and it is a composite benchmark rather than a single test.
The score reflects the model’s performance on specific capabilities and industries. The index includes evaluations of agentic knowledge work, agentic real-world work tasks, agentic SaaS workflows, agentic coding and terminal use, reasoning and knowledge, professional document reasoning, physics reasoning, and more. It also measures agentic business operations, quantitative analysis on spreadsheets and documents, agentic tool use, and Kubernetes incident root-cause analysis.
The index also tests visual reasoning and medical long context reasoning. It includes the AA-Briefcase Elo and the AA-Omniscience Index, which measure intelligence across a range of industries.
What is notable about the index is that it is a composite. It does not measure agentic knowledge work, agentic real-world work tasks, agentic SaaS workflows, agentic coding and terminal use, reasoning and knowledge, professional document reasoning, physics reasoning, or any other single capability in isolation. It measures a model’s performance across all of them at once.
That is the point. The index is not a test of a single skill. It is a test of a model’s general intelligence, and a score of 46 is well above the median.
The Price Breakdown
The pricing is where the story gets interesting. MiMo-V2.6-Pro costs $0.43 per 1M input tokens and $0.87 per 1M output tokens. That is somewhat expensive for the input rate and moderately priced for the output rate, compared to models of a similar size. The blended 7:2:1 cache hit/input/output ratio puts the price at $0.18 per 1M tokens.
The model is slower than average, with a throughput of 54 tokens per second. It also takes over 2 seconds to produce its first token. That latency adds up when you are generating 140M output tokens.
The output volume is also a problem. A model that generates 140M output tokens is going to take a long time to finish, no matter how fast it is thinking. The end-to-end response time is calculated based on the time to first token, the “thinking” time for reasoning models, and the output speed.
The Design Tradeoffs
MiMo-V2.6-Pro is a Mixture of Experts (MoE) model with 1.0 trillion total parameters, but only 42 billion active parameters are used during inference. The model has a 1M tokens context window, which is large and gives it room to handle long conversations and document history. It is open weights, meaning its public weights can be downloaded and self-hosted, and it is released under the MIT license, allowing commercial use.
The model supports image input and can analyze, describe, and answer questions about images. It is multimodal, processing text, image, speech, and video input and generating text output. The model is available through one API provider.
“MiMo-V2.6-Pro is amongst the leading models in intelligence and reasonably priced when comparing to other open weight models of similar size. It’s also slower than average and somewhat verbose.”
That is the honest assessment of the model. It is smart and affordable, but it is slow and chatty. Those are not necessarily bad traits. A model that is smart is useful. A model that is affordable is accessible. But a model that is slow and chatty is frustrating.
How the Comparisons Work
The intelligence and performance analysis compares MiMo-V2.6-Pro against models of the same class. Metrics are compared against models of the same class, including non-reasoning models, reasoning models, open weights models, and proprietary models.
- Non-reasoning models are compared only with other non-reasoning models.
- Reasoning models are compared across both reasoning and non-reasoning models.
- Open weights models are compared only with other open weights models of the same size class: tiny (≤4B parameters), small (4B–40B parameters), medium (40B–150B parameters), and large (>150B parameters).
- Proprietary models are compared across proprietary and open weights models of the same price range, using a blended 3:1 input/output price ratio: <$0.15 per 1M tokens, $0.15–$1 per 1M tokens, and >$1 per 1M tokens.
The comparisons are based on a number of metrics, including intelligence, output tokens, pricing, speed, and latency. Xiaomi’s API is the source of the data, and artificialanalysis.ai’s analysis draws on comparisons against models of similar size and against proprietary models.
The Verdict on the Model
MiMo-V2.6-Pro is a capable model with real strengths. Its Intelligence Index score of 46 is well above average, and its multimodal support and MIT license make it a practical choice for developers who want to build on open weights. The 140M output tokens it generates are a feature, not a bug, for users who want detailed answers.
But the model’s weaknesses are also real. Its throughput is slower than the median, and its latency is higher than average. The $0.43 per 1M input tokens and $0.87 per 1M output tokens pricing is above the median, which means it is not the cheapest option out there.
The model’s $0.87 per 1M output tokens is moderately priced, which is a relief for users who care more about the answers than the questions. The blended 7:2:1 cache hit/input/output ratio puts the price at $0.18 per 1M tokens, which is a reasonable rate for a model of this size.
The model’s 42 billion active parameters and 1M tokens context window give it the capability to handle complex tasks, but the slow throughput and high output volume mean it is not the fastest model out there. Users who need quick answers should look elsewhere. Users who need detailed answers should consider MiMo-V2.6-Pro.
The model’s MIT license and open weights design mean it is free to use commercially, which is a significant advantage. The company has released MiMo-V2.6-Pro, and the intelligence and performance analysis provides a track record of performance and pricing to consider. That track record shows that Xiaomi is willing to publish its own assessments of its models, which is a useful signal for users who want to know what they are getting.
Source material: “MiMo-v2.6-Pro: Intelligence, Performance and Price Analysis,” artificialanalysis.ai.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.