# Xiaomi MiMo V2.6 Pro vs Flash: benchmarks, pricing and which to use in BrainBox

> Xiaomi's MiMo V2.6 Pro and Flash, MIT open weights with a 1M context window, are now in the BrainBox Assistant: benchmarks, Artificial Analysis score, intelligence units per task and which to pick.

- Source: https://www.usebrainbox.com/en/blog/xiaomi-mimo-v2-6-pro-vs-flash-benchmarks-pricing-which-to-use-in-brainbox
- Author: BrainBox Team
- Published: 2026-09-21
- Updated: 2026-09-22
- Language: en
- Tags: models, Xiaomi, product

## Summary

MiMo V2.6 Pro and MiMo V2.6 Flash are Xiaomi's new open weight models, released on September 21, 2026 under the MIT license, and both can be selected in BrainBox with a one million token context window and image input. Pro, with 1.02 trillion parameters and 42 billion active, scores 46 on the Artificial Analysis Intelligence Index, the highest of the 114 open weight models measured. Flash, with 309 billion and 15 billion active, stays close on coding and agent tests. In BrainBox they use few intelligence units: a case file review is about 7 units with Pro and 4 with Flash, and a 450 page case file read whole in a single pass about 14 with Pro and 6 with Flash. Both require a plan with full model access.

---
MiMo V2.6 Pro and MiMo V2.6 Flash, the models Xiaomi released on September 21, 2026 with open weights under the [MIT license](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL), can now be selected in the BrainBox Assistant. Both have a one million token context window and accept images. Pro is the highest scoring open weight model on the [Artificial Analysis Intelligence Index](https://artificialanalysis.ai/models/mimo-v2-6-pro) as of today; Flash stays close on the agent tests and uses a third of the intelligence units. Both require a plan with full model access.

> **Remember**
>
> One million tokens of context for a handful of intelligence units changes one concrete decision: asking the Assistant to read a whole case file in a single pass stops being expensive. A 450 page case file is about 14 units with Pro and 6 with Flash. Page citations stay the same.

## What are MiMo V2.6 Pro and Flash?

They are two mixture of experts models, or MoE, from Xiaomi's MiMo team, trained in a single reinforcement learning run that mixes thousands of task environments: programming, general agents, visual tasks and cybersecurity, according to [the technical report](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL/blob/main/MiMo_V2_6_technical_report.pdf). Xiaomi published the weights on Hugging Face and ModelScope under the MIT license: [MiMo-V2.6-Pro-RL](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL) and [MiMo-V2.6-Flash-RL](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL). Alongside the weights it published the technical report, the training environments, the RL framework and a distilled 9 billion parameter model, MiMo-V2.6-Distill-Qwen-9B. The figures in the table come from [the model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL) and [Xiaomi's model page](https://mimo.mi.com/models/en-US/mimo-v2.6-pro).

| | MiMo V2.6 Pro | MiMo V2.6 Flash |
|---|---|---|
| Total parameters | 1.02 trillion | 309 billion |
| Active parameters per token | 42 billion | 15 billion |
| Context | 1M tokens | 1M tokens |
| Maximum output | 128,000 tokens | 128,000 tokens |
| Inputs | Text, image, audio, video | Text, image, audio, video |
| License | MIT, open weights | MIT, open weights |
| BrainBox picker category | Pro | Default |

In BrainBox, audio and video are not sent to the model: they are transcribed at indexing time and the model receives the transcript with timestamps. The multimodal input it does use is images.

## What do the benchmarks say?

Artificial Analysis measures models on its own. MiMo V2.6 Pro scores [46](https://artificialanalysis.ai/models/mimo-v2-6-pro) on the Intelligence Index v4.3, first among the 114 open weight models it has measured, and the same score as Grok 4.7, released the same day. According to [the leaderboard](https://artificialanalysis.ai/leaderboards/models), it sits above Grok 4.6 at 44, Gemini 3.8 Flash at 41 and DeepSeek V4.1 Flash at 39, and seven points below Claude Fable 5.1 and GPT-6 Astra, which score 53 in their maximum setting. It generates about 125 tokens per second with 2.19 seconds to first token, and Artificial Analysis calls it "somewhat verbose". Flash has no published score yet.

The agent benchmarks Xiaomi publishes on [the model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL) are the company's own, including the numbers for the closed models, and should be read with that in mind:

| Benchmark, as reported by Xiaomi | V2.5 Pro | V2.6 Pro | V2.6 Flash | Claude Opus 5 |
|---|---|---|---|---|
| DeepSWE v1.1, software agent | 19.0 | 71.9 | 67.9 | 74.0 |
| ProgramBench | 12.5 | 26.5 | 26.0 | 37.0 |
| Terminal Bench 4.0 | 1.5 | 34.9 | 28.8 | 49.0 |
| Terminal Bench 2.1 | 65.2 | 89.9 | 87.6 | 89.1 |
| Toolathlon-Verified, tool use | 49.1 | 76.9 | 73.6 | 80.6 |
| AutomationBench v1.0.6 | 16.0 | 53.1 | 52.3 | 50.3 |
| CyberGym, cybersecurity | 40.0 | 94.0 | 95.1 | |
| ExploitBench | 16.6 | 47.9 | 25.3 | 70.0 |

Xiaomi's [technical report](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL/blob/main/MiMo_V2_6_technical_report.pdf) describes "performance comparable to that of frontier models across various domains", and its own table shows where it is not: Claude Opus 5 stays ahead on DeepSWE v1.1, ProgramBench and Terminal Bench 4.0, and the gap on ExploitBench is wide. Pro belongs to the frontier group for agent tasks without leading it, and in BrainBox uses a fraction of the intelligence units Grok 4.7 uses on the same task. Flash stays within 1 to 4 points of Pro on coding and general agent tests; on offensive cybersecurity the distance is larger.

## How does the BrainBox Assistant work with MiMo V2.6?

The BrainBox Assistant is an agent. You choose which Boxes and files it may use and which model does the reasoning; the agent decides the steps. Each one shows in the chat: "Listing sources", "Searching documents", "Reading document" with the pages it opened, "Saving note" when it sets something aside for later, and at the end the answer with page citations. With a million tokens of context, MiMo V2.6 can hold many more reads in one conversation before the agent has to compress what it read, and it can read a document of hundreds of pages whole when the task calls for it.

## How do you pick them in BrainBox?

  - **Check that your plan has full model access** Xiaomi's models carry the Full plan badge. They appear on paid personal accounts; in a workspace, the admin enables them in the AI policy.
  - **Open the Assistant and choose the sources** In the sidebar you tick the Boxes and files the agent may use. The model picker sits in the bottom bar of the chat and is searchable.
  - **Type 'MiMo'** MiMo V2.6 Pro, MiMo V2.6 Flash and the V2.5 versions appear with their context and badges.
  - **Choose Pro or Flash for the task** Pro to review and reason over a whole case file; Flash for frequent questions and batch summaries.
  - **Ask for the task** If you want a document read whole, say so in the message. Otherwise the agent searches and reads the pages it needs. The answer still carries page citations.

*Simulation of the BrainBox Assistant model picker with Xiaomi's two models.*

BrainBox is an AI workspace for documents that answers with citations to the exact page. With MiMo V2.6 the answer comes from an open model, and the citation still points to the document in the Box.

## How many intelligence units does each task use?

BrainBox does not measure usage in tokens but in intelligence units, and each task uses a different number depending on how many searches and reads the agent makes, how much the model reasons and how much it writes. Estimates for the standard plan, with Grok 4.7 alongside for reference:

| Task in the Assistant | MiMo V2.6 Pro | MiMo V2.6 Flash | Grok 4.7 |
|---|---|---|---|
| Single cited question, one or two searches | ≈ 1 unit | ≈ 1 unit | ≈ 3 units |
| Spreadsheet analysis with the code interpreter | ≈ 3 units | ≈ 2 units | ≈ 8 units |
| Case file review with about ten searches and reads | ≈ 7 units | ≈ 4 units | ≈ 20 units |
| Full read of a 200 page contract, page by page | ≈ 7 units | ≈ 3 units | ≈ 22 units |
| PDF report built from a case file, with code and about fifteen steps | ≈ 11 units | ≈ 5 units | ≈ 35 units |
| A 450 page case file read whole in a single pass | ≈ 14 units | ≈ 6 units | ≈ 95 units |

These are estimates: a longer answer, more sources or more agent steps push the number up. For reference, the Elite plan includes 400 units a month; with Flash, a case file review uses little more than one question with Grok 4.7.

## What happens if the Box is larger than the million token window?

Nothing is left out. The Assistant does not load the whole Box into the model: it searches the selected sources and reads the pages that answer the question, each with its file and page. It works the same for a Box of 200 pages or 20,000. What changes with MiMo V2.6 is how much it can hold as read at once: over a thousand pages before it has to compress what it read into notes.

## When does Pro make sense, when Flash, and when neither?

The advantage of these models in BrainBox is context: a million tokens with no usage cliff. With Grok 4.7 a single step over about 300 pages nearly doubles the units it uses; with Xiaomi's models usage grows evenly up to the limit.

| Task in the Assistant | Model | Why |
|---|---|---|
| [Reading a whole contract without skipping clauses](/en/blog/compare-two-contract-versions-with-ai) in a single pass | Pro | 1M context with no usage cliff; reasons better than Flash on long tasks |
| Summarizing 50 meeting minutes or sorting emails in batches | Flash | One unit per question; the quality gap barely shows on short tasks |
| Single cited question | Flash or the default model | One unit either way; the model matters little |
| [PDF report](/en/blog/ask-ai-for-a-report-and-get-a-designed-pdf-or-powerpoint) or data analysis with the code interpreter | Pro | Its best results are on coding and tool use agents |
| Opinion or review where the error margin must be minimal | Claude or GPT | Still [seven points ahead](https://artificialanalysis.ai/leaderboards/models) on the independent index; the extra units are justified |
| Teams with a restricted provider policy | Depends on the policy | The admin controls which providers and models are allowed |

> **Ask both the same question**
>
> Switching models in the picker does not restart the chat. Ask with Flash, switch to Pro and compare. If both answers cite the same pages and say the same thing, Flash is enough for that task.

## How is this different from using MiMo in Xiaomi's app?

Xiaomi offers the models in its AI Studio, MiMo Code, MiMo Desktop and its API, [according to the model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL). There you work with what you paste into the conversation. In BrainBox the model works on the Box.

| | Xiaomi AI Studio | BrainBox with MiMo V2.6 |
|---|---|---|
| Sources | What you upload in that conversation | PDF, Office, transcribed audio and images already indexed, with per Box permissions |
| Citations | Do not point to a page in your files | Page, passage and expanded context per claim |
| Switching models | Xiaomi models only | Pro and Flash next to Claude, GPT, Gemini and Grok, in the same chat |
| Team | Individual account | Shared Box with roles and a per workspace model policy |
| Usage | Depends on the platform | Intelligence units per use, based on the chosen model |

A [research team](/en/solutions/research) can upload a hundred papers, pick Flash for exploratory questions and move to Pro for the final analysis, [checking each citation](/en/blog/how-to-verify-an-ai-answer-about-your-documents) in the same panel. The rest of the [document workspace](/en/blog/brainbox-2026-workspace-for-document-work) does not change with the model.

According to [the technical report](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL/blob/main/MiMo_V2_6_technical_report.pdf), the reinforcement learning phase cost $2.6 million for Pro and $0.9 million for Flash, with batches of 1,568 prompts and sequences of 110,000 to 150,000 tokens. In BrainBox, reading a 450 page case file whole in a single pass uses about 14 intelligence units with Pro and 6 with Flash.

## Sources

- Xiaomi MiMo, model cards for [MiMo-V2.6-Pro-RL](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL) and [MiMo-V2.6-Flash-RL](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL) on Hugging Face: parameters, context, modalities, MIT license, benchmark table and availability.
- Xiaomi MiMo, [MiMo-V2.6 technical report](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL/blob/main/MiMo_V2_6_technical_report.pdf): reinforcement learning, RL cost, distilled model and open suite.
- Xiaomi, [MiMo-V2.6-Pro model page](https://mimo.mi.com/models/en-US/mimo-v2.6-pro): 128K maximum output and modalities.
- Artificial Analysis, [MiMo-V2.6-Pro](https://artificialanalysis.ai/models/mimo-v2-6-pro): Intelligence Index, rank among open weight models, speed and latency.
- Artificial Analysis, [model leaderboard](https://artificialanalysis.ai/leaderboards/models): scores for Grok 4.7, Grok 4.6, Gemini 3.8 Flash, DeepSeek V4.1 Flash, Claude Fable 5.1, GPT-6 Astra and MiMo-V2.5-Pro.

## FAQ

### What is the difference between MiMo V2.6 Pro and Flash?

Size and usage. Pro has 1.02 trillion total parameters with 42 billion active per token; Flash has 309 billion total with 15 billion active. Both have a 1 million token context window and accept images. On Xiaomi's published benchmarks, Flash lands 1 to 4 points below Pro on coding and general agent tests, and much further behind on offensive cybersecurity: on ExploitBench, 25.3 against 47.9. In BrainBox, Flash uses about a third of the intelligence units Pro uses.

### What do they cost in BrainBox?

Usage is measured in intelligence units. Estimates on the standard plan: a single cited question, 1 unit with either model; a case file review with searches and page reads, about 7 units with Pro and 4 with Flash; a PDF report built from a case file, about 11 with Pro and 5 with Flash; a 450 page case file read whole in a single pass, about 14 with Pro and 6 with Flash. There is no long context usage cliff.

### Why do I need a plan with full model access?

Xiaomi's models carry the Full plan badge in the picker. On personal accounts they appear on paid plans; in a workspace, the admin decides in the AI policy whether the team can use them.

### How good is MiMo V2.6 Pro compared with Claude or GPT?

Artificial Analysis gives it 46 on its Intelligence Index v4.3, the same score as Grok 4.7 and above Grok 4.6 at 44, Gemini 3.8 Flash at 41 and DeepSeek V4.1 Flash at 39. Claude Fable 5.1 and GPT-6 Astra score 53. In Xiaomi's own table, Claude Opus 5 stays ahead on DeepSWE v1.1, ProgramBench and Terminal Bench 4.0.

### What if my Box is larger than the million token window?

Nothing is left out. The Assistant does not load the whole Box at once: it searches the sources you chose and reads the pages that answer the question, each with its citation. The size of the Box does not limit which model you can use.

### Do my documents go to Xiaomi's servers?

Your documents stay stored in BrainBox. What the model receives is the context passage for each question, as with any other model in the picker. If your compliance policy excludes certain providers, the admin can block them in the workspace.