MiMo-V2.6 Guide: VRAM Requirements, GGUF Builds

September 27, 2026November 12, 2026

About This Model

MiMo-V2.6 is an open-weight model series from Xiaomi’s MiMo team. It is an omni-modal MoE (Mixture of Experts) family that handles text, images, video and audio in a single model, with a context window of up to one million tokens. The license is MIT.

The series has two models of different sizes.

Both interleave sliding-window attention (SWA, 128-token window) with global attention, and both include a 681M-parameter vision encoder (MiMo ViT) and audio encoders. A 5-layer MTP drafter for speculative decoding is built in and predicts the next 7 tokens per forward pass.

As the “RL" in the names suggests, the theme of the series is scaling reinforcement learning (RL). Instead of training coding, general agents, visual tasks and cybersecurity separately, it trains them together in one mixed RL run (the model card calls this “You Only RL Once").

What Makes It Stand Out

1. Agentic performance jumped by an order of magnitude over the previous generation. An excerpt from the publisher’s comparison table (bold marks the best value in each row):

• The biggest gains over MiMo-V2.5 Pro are on tasks that require working through many steps to the end. DeepSWE (long-horizon development) rose from 19.0 to 71.9, AutomationBench (orchestrating business workflows across REST APIs) from 16.0 to 53.1, and Terminal Bench 4.0 from 1.5 to 34.9.
• Despite being open-weight, it matches top commercial models on some rows. It has the highest score in the table on Terminal Bench 2.1 and AutomationBench, and ties Claude Opus 5 on Agents’ Last Exam.

2. Xiaomi published how it was trained, not just the weights. Besides the technical report, Xiaomi released the RL environments used in training as a dataset (coding, cyber, knowledge work, web development and music composition), its training framework (a fork of verl), and a small 9B model for reproducing the experiments on those environments (MiMo-V2.6-Distill-Qwen-9B). The technical report also documents failures, such as expert load collapsing onto a few experts when the MoE router was left trainable during RL, and training interruptions caused by GPU memory errors and network outages.

3. The smaller Flash comes close to Pro on most rows. On DeepSWE, Terminal Bench 2.1 and OSWorld-Verified, the gap to Pro is within 4 points. With less than a third of Pro’s total parameters, Flash delivers most of the performance for coding and business-agent use.

Caveats and Weaknesses

• It does not reach the top commercial models on tasks that require analyzing and rebuilding or breaking into unfamiliar targets. On ProgramBench (rebuilding a program from its binary and documentation) it is more than 10 points below Claude Opus 5, and on Terminal Bench 4.0 it scores 34.9 against 49.0. On ExploitBench (actually exploiting vulnerabilities) it falls well short of the three commercial models (70.0–78.5). On OSWorld-Verified (operating a computer through its screen) it is also below all three.
• Pro and Flash diverge sharply on offensive security tasks. ExploitBench is 47.9 vs 25.3 and ExploitGym 17.8 vs 6.0, so Flash is not a substitute for Pro in this area.
• Some benchmarks are internal. MiMo Code Bench, MiMo Cyber Bench and MiMo VisualCoding in the model card are Xiaomi’s in-house evaluations that third parties cannot reproduce under the same conditions. They are left out of the excerpt above.

Running It Locally

• The publisher provides Transformers-format weights and recommends SGLang and vLLM. The Pro launch example uses tensor parallelism 16 across 2 nodes, well beyond personal hardware. The Flash examples also assume multiple GPUs (tensor parallelism 4–8).
• For llama.cpp, the entry point is ggml-org’s GGUF build of Flash-RL. It offers an MXFP4 build and a build with part of the experts in Q2_K, plus MTP and DFlash sidecars for speculative decoding and an mmproj for the vision and audio encoders, all in the same repository. See the memory table further down this page for how much memory each configuration needs. The model card notes that the Q2_K build was made without an imatrix.
• Recommended sampling is temperature=1.0 and top_p=0.95. With vLLM, add --reasoning-parser mimo and --tool-call-parser mimo. Because stable vLLM may lag, a dedicated Docker image built for MiMo-V2.5 is recommended.
• The license is MIT, and commercial use is allowed.

Sources: the model cards of XiaomiMiMo/MiMo-V2.6-Pro-RL, XiaomiMiMo/MiMo-V2.6-Flash-RL and ggml-org/MiMo-V2.6-Flash-RL-GGUF, and the technical report “MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement" in the same repository (as of 2026-09-27). Benchmark figures are as published by the developer.

Our Coverage and Data

Everything Local Model Watch has published about the MiMo-V2.6 family: 4 article(s) covering the base model and its fine-tunes, plus converted builds we tracked after publication. Memory requirements below are computed by this site from file sizes, not quoted from model cards. Part of our model family index.

At a Glance

Hardware Requirements

Estimated requirements (calculated by Local Model Watch)

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. File sizes are measured from the converted build DevQuasar/XiaomiMiMo.MiMo-V2.6-Pro-RL-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Figures are for XiaomiMiMo/MiMo-V2.6-Pro-RL. Other sizes in this family are covered in the articles below.

Can You Run It Locally?

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

The publisher ships safetensors, but DevQuasar/XiaomiMiMo.MiMo-V2.6-Pro-RL-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.

This assessment is for XiaomiMiMo/MiMo-V2.6-Pro-RL.

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
• Available builds in ggml-org/MiMo-V2.6-Pro-RL-GGUF: Q2_K 390.0GB / MXFP4 516.1GB
• Available builds in DevQuasar/XiaomiMiMo.MiMo-V2.6-Pro-RL-GGUF: Q2_K 347.8GB / Q5_K_M 455.7GB / MXFP4 516.1GB / Q4_K_M 575.4GB
• Available builds in mlx-community/MiMo-V2.6-Pro-RL-mxfp4-q8: MLX 4bit 515.8GB

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.

Popularity on Hugging Face

Weekly snapshots recorded by Local Model Watch after publication. “Downloads" is Hugging Face’s own trailing-30-day count, not a lifetime total.

Articles (the family’s own models first, then newest)

Repositories

• XiaomiMiMo/MiMo-V2.6-Flash-MOPD
• XiaomiMiMo/MiMo-V2.6-Pro-RL
• XiaomiMiMo/MiMo-V2.6-Flash-RL
• XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B
• ggml-org/MiMo-V2.6-Flash-RL-GGUF
• ggml-org/MiMo-V2.6-Distill-Qwen-9B-GGUF

Last updated 2026-11-12 (JST). The explanation at the top of this page was written with the help of AI from the primary sources it cites. The tables and lists under “Our Coverage and Data" are assembled by code from our article log.