Xiaomi Releases MiMo-V2.6-Flash-MOPD and Pro-MOPD Models

September 28, 2026October 9, 2026

At a Glance

Values determined by this site’s code when the information was collected. Dates are JST.

What we checked ourselves

• The same Japanese text takes about as many tokens as with the Qwen3 tokenizer.

Details and conditions are in “Our Own Measurements” below.

Overview

Xiaomi has released new sparse MoE (Mixture of Experts) multimodal models, MiMo-V2.6-Flash-MOPD and MiMo-V2.6-Pro-MOPD. These models apply an upgrade called MOPD2 (Multi-teacher On-policy Distillation 2) to the checkpoints of the previously released “MiMo-V2.6-Flash-RL" and “MiMo-V2.6-Pro-RL".

MOPD2 is a method that distills multiple domain-specific teacher models into a student model on-policy. These include “mixRL teachers" trained on verifiable tasks and “SFT teachers" trained on synthetic demonstrations for open-domain tasks where reliable reward design is difficult. This expands the model’s ability to adapt to domains that are difficult to verify during training, such as long-term game development, scientific research, and embodied intelligence. Furthermore, it was developed to effectively diagnose and mitigate an easily overlooked failure mode in agent environments known as “tool-call repetition," where the model repeatedly makes the same tool calls and stalls progress.

Specifications

The main specifications of the two released models are as follows.

MiMo-V2.6-Flash-MOPD

• Architecture: Sparse MoE (Mixture of Experts)
• Total Parameters: 309B
• Active Parameters: 15B
• Context Length: 1M tokens
• Supported Modalities: Text, image, video, audio
• Vision Encoder: 681M parameter MiMo ViT (28 layers: 24 SWA + 4 Full)
• Audio Encoder: 308M AudioTokenizer + 127M audio patch encoder
• Multi-Token Prediction (MTP): 5-layer speculative decoder (DFlash style, predicting 7 tokens per forward pass)

The configuration details for the LLM backbone and vision encoder of “MiMo-V2.6-Flash-MOPD" are as follows.

MiMo-V2.6-Pro-MOPD

• Architecture: Sparse MoE (Mixture of Experts)
• Total Parameters: 1.02T
• Active Parameters: 42B
• Context Length: 1M tokens
• Supported Modalities: Text, image, video, audio
• Vision Encoder: 681M parameter MiMo ViT (28 layers: 24 SWA + 4 Full)
• Audio Encoder: 308M AudioTokenizer + 127M audio patch encoder
• Multi-Token Prediction (MTP): 5-layer speculative decoder

Performance

Specific numerical score tables for general benchmark suites or empirical benchmark data from Hugging Face are not provided for this model group. However, based on reports from the development team at Xiaomi MiMo, specific performance improvements in agent tasks are qualitatively indicated.

According to the development team’s analysis, the most prominent issue that emerged after the initial release of “MiMo-V2.6" was the “tool-call repetition" problem in agent environments. This is a failure mode where the model repeatedly issues the same or very similar tool calls over and over, wasting time and context while making zero progress on the task itself. Because this phenomenon does not explicitly fail as an error, it tends to be easily overlooked.

It is reported that applying the MOPD (Multi-teacher On-policy Distillation) stage has efficiently solved this issue. Specifically, they adopted a lightweight modification approach involving short-term specialized teacher training integrated into the regular MOPD pass. As a result, compared to models from the RL (Reinforcement Learning) stage, the tool-call repetition rate at the response level is significantly reduced across various context lengths and agent evaluation frameworks (agent harnesses), improving operational stability in practical agent use.

Strengths and Use Cases

MiMo-V2.6-Flash-MOPD and MiMo-V2.6-Pro-MOPD possess omni-modal processing capabilities that comprehensively handle text, images, video, and audio, along with strengths specialized for advanced autonomous agent tasks.

Specifically, they are suited for the following fields and applications:

Agent Systems and Autonomous Tool Execution

Through this MOPD2 distillation, the issue of endlessly repeating similar tool calls without progress in agent environments is suppressed. This is expected to enable stable execution without getting stuck in long-term tasks that require continuously calling external tools and APIs to achieve goals autonomously (such as game development, scientific research workflow automation, and embodied intelligence). Additionally, the model card specifies the use of dedicated parsers for reasoning and tool calling (--reasoning-parser mimo, --tool-call-parser mimo), making it optimized for structured agent control.

Multimodal Understanding Leveraging Ultra-Long Context (1M Tokens)

Equipped with a large-scale context window of up to 1M tokens, the models are capable of analyzing lengthy documents as well as long-form video understanding and continuous audio processing. Since video and image recognition via the vision encoder (MiMo ViT) and speech recognition via the AudioTokenizer and audio patch encoder are integrated within the same model, it supports complex analysis and conversational tasks in mixed video and audio media.

Multilingual Conversations

English (en) and Chinese (zh) are provided as supported language tags, anticipating operation as advanced conversational systems or multilingual agents based on these two languages.

How to Get It

Both models are released under the MIT License and are not gated models requiring agreement to terms of use, allowing anyone to freely download and use them. They are distributed in FP8 (8-bit) safetensors format.

The repositories are available on Hugging Face and ModelScope.

• Hugging Face:
• XiaomiMiMo/MiMo-V2.6-Flash-MOPD – XiaomiMiMo/MiMo-V2.6-Pro-MOPD
• ModelScope:
• XiaomiMiMo/MiMo-V2.6-Flash-MOPD – XiaomiMiMo/MiMo-V2.6-Pro-MOPD

They are also noted as available via cloud services and GUI environments such as AI Studio, MiMo Code, Xiaomi MiMo Desktop, Xiaomi MiMo Open Platform API, and OpenRouter.

Serving instructions using SGLang and vLLM are published as inference engines. Recommended sampling parameters are temperature=1.0 and top_p=0.95.

Deployment with SGLang

Using the official Docker image lmsysorg/sglang:latest is recommended. EAGLE can be set as the speculative decoding algorithm.

Startup example for MiMo-V2.6-Flash-MOPD:

sglang serve \
  --trust-remote-code \
  --model-path XiaomiMiMo/MiMo-V2.6-Flash-MOPD \
  --tp 8 \
  --dp 2 \
  --enable-dp-attention \
  --enable-dp-lm-head \
  --mm-enable-dp-encoder \
  --mem-fraction-static 0.65 \
  --chunked-prefill-size 16384 \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4 \
  --enable-multi-layer-eagle \
  --reasoning-parser mimo \
  --tool-call-parser mimo \
  --host 0.0.0.0 \
  --port 30000

Startup example for MiMo-V2.6-Pro-MOPD (2-node configuration):

sglang serve \
  --trust-remote-code \
  --model-path XiaomiMiMo/MiMo-V2.6-Pro-MOPD \
  --tp 16 \
  --dp 2 \
  --enable-dp-attention \
  --mm-enable-dp-encoder \
  --ep 16 \
  --moe-a2a-backend deepep \
  --moe-dense-tp-size 1 \
  --mem-fraction-static 0.7 \
  --max-running-requests 128 \
  --chunked-prefill-size 32768 \
  --page-size 64 \
  --swa-full-tokens-ratio 0.3 \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4 \
  --enable-multi-layer-eagle \
  --reasoning-parser mimo \
  --tool-call-parser mimo \
  --host 0.0.0.0 \
  --port 30000 \
  --nnodes 2 \
  --node-rank <node-rank> \
  --dist-init-addr <node0-ip>:20000

Deployment with vLLM

Since stable vLLM releases may have delayed support, the pre-built Docker image vllm/vllm-openai:mimov25-cu129 is recommended.

Startup example for MiMo-V2.6-Flash-MOPD:

vllm serve XiaomiMiMo/MiMo-V2.6-Flash-MOPD \
  --tensor-parallel-size 4 \
  --trust-remote-code \
  --gpu-memory-utilization 0.95 \
  --max-model-len auto \
  --reasoning-parser mimo \
  --tool-call-parser mimo \
  --enable-auto-tool-choice \
  --generation-config vllm

Startup example for MiMo-V2.6-Pro-MOPD:

vllm serve XiaomiMiMo/MiMo-V2.6-Pro-MOPD \
  --tensor-parallel-size 8 \
  --trust-remote-code \
  --gpu-memory-utilization 0.95 \
  --max-model-len auto \
  --reasoning-parser mimo \
  --tool-call-parser mimo \
  --enable-auto-tool-choice \
  --generation-config vllm

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp via a converted build.

The publisher ships safetensors, but ggml-org/MiMo-V2.6-Pro-MOPD-GGUF provides a GGUF build you can use.

License — mit (Commercial use allowed): Permits commercial use, modification and redistribution, provided the copyright notice and license text are retained.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Values we measured ourselves on our server (no GPU) by actually reading and running this model’s files — not figures copied from the model card. How we measure, and the results for every model: Our Measurements.

Japanese Token Efficiency

The same Japanese text takes about as many tokens as with the Qwen3 tokenizer.

Counted with the tokenizer.json of XiaomiMiMo/MiMo-V2.6-Flash-MOPD on a fixed text we wrote ourselves (876 Japanese characters across news, conversation, technical docs, a formal email, travel writing and a recipe) and its English translation. Fewer tokens mean more Japanese fits in the context window.

Inside the GGUF File

File: MiMo-V2.6-Pro-MOPD-Q2_K-00001-of-00007.gguf (390.01GB, Q2_K). We read only the header (metadata) of the file, not the weights.

The quant name in the file name describes the file as a whole; in practice layers mix several types. The average bits per weight is the measured file data divided by the number of weights.

Quantized and Converted Variants

File sizes of each build:

• Available builds in ggml-org/MiMo-V2.6-Pro-MOPD-GGUF: Q2_K 390.0GB / MXFP4 516.1GB

In addition, 22 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

• MiMo-V2.6-Pro-RL Text Generation Model: ~417GB Memory, GGUF Builds

What to Read Next

• Explore the same model family → MiMo-V2.6 family overview (4 articles, 4 converted builds)
• Engines that run this model → llama.cpp / Ollama / vLLM
• Formats this model is available in → GGUF format guide and models
• Learn about the publisher → Xiaomi (MiMo): models, licenses and articles
• Other models for the same task → Other text generation models

Sources

• XiaomiMiMo/MiMo-V2.6-Flash-MOPD (Hugging Face)
• XiaomiMiMo/MiMo-V2.6-Pro-MOPD (Hugging Face)

Update History

• 2026-09-28: Added our own measurements: Japanese token efficiency.
• 2026-10-03: Added converted builds to “Quantized and Converted Variants”: ggml-org/MiMo-V2.6-Pro-MOPD-GGUF
• 2026-10-06: Added our own measurements: what is inside the GGUF file.