MiMo-V2.6-Flash-RL-GGUF Multimodal MoE Model: ~141GB Memory
At a Glance
| Item | Value |
|---|---|
| Repository | ggml-org/MiMo-V2.6-Flash-RL-GGUF |
| Family guide | MiMo-V2.6 guide (4 articles) |
| Publisher guide | Xiaomi (MiMo): models and licenses |
| Published | 2026-09-22 |
| License | mit |
| Formats | GGUF |
| Source type | Primary source (the publisher itself) |
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
ggml-org has released “ggml-org/MiMo-V2.6-Flash-RL-GGUF", a GGUF quantized version of the multimodal model “MiMo-V2.6-Flash-RL" developed by XiaomiMiMo. Created using automated conversion tools, this model includes quantization formats such as MXFP4 and Q2_K, alongside MTP sidecars (MXFP4 and Q8_0) for utilizing speculative decoding features (--mtp), as well as a Q8_0 mmproj for image and audio encoders.
The original model, “XiaomiMiMo/MiMo-V2.6-Flash-RL“, was built to scale reinforcement learning (RL) for self-improvement. It adopts a native omni-modal configuration that handles text, images, video, and audio within a single model, complexly enhancing coding, general agents, visual tasks, and cybersecurity through a single RL run.
Specifications
- Parameters: 309B total / 15B active parameters
- Architecture: Sparse MoE (Mixture of Experts)
- Context Length: 1M tokens
Performance
The evaluation results below are taken from the model card of the original model “XiaomiMiMo/MiMo-V2.6-Flash-RL". Since this model is a GGUF quantized version, slight accuracy variations may occur due to quantization. Note that due to the high number of columns in the table, it has been narrowed down to focus on the original model and major comparative models (numerical values are kept as in the original text).
→ Scroll horizontally to see all columns
| Benchmark | MiMo-V2.6 Flash | MiMo-V2.6 Pro | MiMo-V2.5 Pro | Claude Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Code Agent | |||||
| DeepSWE v1.1 | 67.9 | 71.9 | 19.0 | 74.0 | 73.0 |
| ProgramBench | 26.0 | 26.5 | 12.5 | 37.0 | 25.0 |
| MiMo Code Bench | 61.2 | 63.2 | 40.4 | 68.6 | 59.3 |
| General Agent | |||||
| AutomationBench v1.0.6 | 52.3 | 53.1 | 16.0 | 50.3 | 45.8 |
| Toolathlon-Verified | 73.6 | 76.9 | 49.1 | 80.6 | 74.9 |
| Agents’ Last Exam | 27.6 | 31.6 | 13.2 | 31.6 | 30.8 |
| Terminal Bench 4.0 | 28.8 | 34.9 | 1.5 | 49.0 | 39.9 |
| Terminal Bench 2.1 | 87.6 | 89.9 | 65.2 | 89.1 | 88.8 |
| OSWorld-Verified | 80.8 | 82.0 | – | 83.4 | 83.0 |
| JobBench | 61.2 | 62.0 | 25.0 | 65.7 | 45.4 |
| Cybersecurity | |||||
| CyberGym | 95.1 | 94.0 | 40.0 | – | – |
| MiMo Cyber Bench | 77.2 | 80.2 | 0.0 | – | – |
| ExploitGym | 6.0 | 17.8 | 0.2 | 22.1 | 30.3 |
| ExploitBench | 25.3 | 47.9 | 16.6 | 70.0 | 78.5 |
| SEC Bench Pro | 47.5 | 66.3 | 17.7 | – | 79.1 |
| Visual Agent | |||||
| MiMo VisualCoding | 71.5 | 72.3 | – | 70.0 | 73.4 |
According to measurements by the publishers, MiMo-V2.6-Flash-RL scored 95.1 on CyberGym and 52.3 on AutomationBench v1.0.6, showing results that surpass other models in certain areas. It also achieved high-level scores of 87.6 on Terminal Bench 2.1, which measures whether commands can be actually executed in a terminal to complete tasks to the end, and 80.8 on OSWorld-Verified, which measures the ability to operate actual OS screens to complete tasks. On the other hand, it falls short of other large frontier models in cybersecurity fields such as ExploitBench (25.3) and ExploitGym (6.0), as well as Terminal Bench 4.0 (28.8), clearly dividing its strengths and weaknesses.
Strengths and Use Cases
The areas of strength and intended use cases for this model are based on the capabilities and training design of the original model, “XiaomiMiMo/MiMo-V2.6-Flash-RL".
-
Long-Context Processing with Integrated Multimodality
It supports native omni-modality, handling text, images, video, and audio within a single model. Furthermore, leveraging a context length of up to 1M tokens, it excels at analyzing vast software repositories as a whole, understanding lengthy tool execution histories (tool traces), or processing multi-session dialogues and agent operation logs without information loss. -
Autonomous Agent Operation and Environmental Adaptability
Through a batch reinforcement learning method called “You Only RL Once", coding, general agents, visual tasks, and cybersecurity are trained in a complex manner within a single batch. This enables flexible adaptation and task execution in unfamiliar environments and tools, such as completing tasks via command execution in a terminal or visually recognizing and operating GUI screens of an OS. -
Fast Inference via Speculative Decoding
This model is equipped with a drafter for multi-token prediction (MTP) (utilizing a 5-layer SWA MTP structure) and supports speculative decoding to predict multiple tokens in a single forward pass. It is suitable for conversational assistants and agent executions where response generation speed and efficiency are required.
The main supported languages are English (en) and Chinese (zh), with expected use cases including conversational tasks, advanced tool calling, and complex multimodal analysis.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| More than 141GB of VRAM (multi-GPU or CPU offload required) | Q2_K | 117.5GB | 141.1GB |
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Runs in Ollama, LM Studio and llama.cpp as-is.
It is distributed in GGUF, so no conversion is needed.
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
| Tokenizer | Tokens per 1,000 Japanese characters | Ratio to the same text in English |
|---|---|---|
| This model | 688 | 1.26× |
| Qwen3 | 688 | 1.26× |
| Llama 3.2 | 744 | 1.36× |
| Gemma 3 | 564 | 1.03× |
| gpt-oss | 795 | 1.45× |
| LLM-jp-3 | 497 | 0.85× |
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-RL 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-Flash-RL-Q2_K-00001-of-00002.gguf (117.55GB, Q2_K). We read only the header (metadata) of the file, not the weights.
| Item | Value |
|---|---|
| Architecture (as named in the GGUF) | mimo2 |
| Maximum trained context length | 1,048,576 tokens |
| Layers | 48 |
| Experts | 8 active out of 256 |
| Vocabulary size | 152,576 |
| Chat template | Included (mentions tool calls, has a thinking switch) |
| imatrix | Not recorded in the file |
| Weight types (share of parameters) | Q2_K 65.4% / MXFP4 32.7% / Q8_0 1.7% / Q6_K 0.2% / other 0.0% |
| Average bits per weight | 3.27 bits |
| Embedding / output layer type | Q8_0 / Q6_K |
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.
How to Get It
ggml-org/MiMo-V2.6-Flash-RL-GGUF is available in GGUF format on Hugging Face. The model is released under the MIT license.
Various quantized files and additional modules tailored to different use cases and environments are provided in the repository:
- MXFP4: A configuration that maintains the routed expert sections at native MXFP4 precision.
- Q2_K: A configuration where expert down projections are kept in MXFP4, and gate and up projections are quantized to Q2_K (Note: imatrix calibration is not applied at this stage).
- MTP sidecars (MXFP4 / Q8_0): Sidecar configurations for utilizing the speculative decoding feature (
--mtp). - Q8_0 mmproj: Components for the image encoder (MiMo ViT) and audio encoder.
Setup and server startup using llama.app can be performed by running the following command:
llama serve -hf ggml-org/MiMo-V2.6-Flash-RL-GGUF
Related Articles
- GLM-5.3-Flash-GGUF Vision-Language Model: ~150GB Memory
- OpenJev-GGUF 27.4B Decision-Making Specialized Model: 24GB+ VRAM
- Testing OpenJev GGUF: A Tiny Model for Runtime Loaders
- MiMo-V2.6-Distill-Qwen-9B-GGUF: Our Test Answers, 12GB+ VRAM
What to Read Next
- Find models by VRAM (This model does not fit even in 80GB; its smallest build needs about 141GB) → VRAM quick reference, including models over 80GB
- Explore the same model family → MiMo-V2.6 family overview (4 articles, 4 converted builds)
- Engines that run this model → llama.cpp / Ollama
- What Q2_K mean and where to get this model → GGUF format guide and models / Quantization and model-format glossary
- Learn about the publisher → Xiaomi (MiMo): models, licenses and articles
Sources
- https://huggingface.co/ggml-org/MiMo-V2.6-Flash-RL-GGUF
- https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL
Update History
- 2026-09-23: Rebuilt the article (details are in the Japanese edition).
- 2026-09-23: Rebuilt the article (details are in the Japanese edition).
- 2026-09-24: Rebuilt the article: Added a “can you run it locally?" verdict from the distribution formats, converted builds, engine registries, license terms and measured compression
- 2026-09-25: Changed the title to show what the article covers (VRAM requirements, file list, etc.).
- 2026-09-26: Changed the title to show what the article covers (VRAM requirements, file list, etc.).
- 2026-09-28: Added our own measurements: Japanese token efficiency.
- 2026-10-06: Added our own measurements: what is inside the GGUF file.