Model library / XiaomiMiMo
MiMo-V2.6-Pro-RL is a 524-billion-parameter sparse MoE model designed for reinforcement-learning-driven self-improvement across text, image, video and audio.
8 of 384 experts active per token
Context length in tokens, as published
Weights at the released precision
Licence declared on the repository
Overview
MiMo-V2.6-Pro-RL is the flagship checkpoint of XiaomiMiMo's MiMo-V2.6 series. The publisher states it is built to scale reinforcement learning toward self-improvement, offering native omnimodal capabilities and a one-million-token context for long-horizon tasks such as tool traces and multi-session agent runs.
In deployment the model contains 524.1 billion parameters with a Mixture-of-Experts architecture of 384 experts, eight active per token. It supports a 1,048,576-token context window, is released under an MIT licence, and the weights are stored in fp8 format.
Specification
| Repository | XiaomiMiMo/MiMo-V2.6-Pro-RL |
| Publisher | XiaomiMiMo |
| Published | 21 September 2026 |
| Architecture | MiMoV2ForCausalLM |
| Model type | mimo_v2 |
| Parameters | 524,121,348,864 |
| Experts | 384 total, 8 active per token |
| Layers | 70 |
| Hidden size | 6,144 |
| Attention heads | 128, 8 key/value heads |
| Context length | 1,048,576 tokens |
| Vocabulary | 152,576 tokens |
| Quantisation | fp8 |
| Licence | mit |
Memory
Parameter count multiplied by the bytes each format uses. The key-value cache comes on top.
| Precision | Weights | Notes |
|---|---|---|
| bf16 | 976.3 GiB | Full precision as released |
| fp8 | 488.1 GiB | 8-bit, near-lossless on most models |
| int4 | 244.1 GiB | 4-bit, smallest footprint |
The cache costs 420 KiB per token at 16-bit, so the full 1,048,576-token context of one request needs about 420.0 GiB. Concurrency multiplies that number, not the weights.
That figure is an upper bound: this model uses sliding-window attention with a 128-token window on some layers, so long requests cache less than this.
Hardware
At the released precision, with 90% of the card given to the server and 2.5 GiB kept for the runtime. The context column is what is left for the cache on that setup.
| GPU | Cards needed | Free for cache | Context it holds |
|---|---|---|---|
| RTX 4090 24GB | more than 8 | — | — |
| RTX 5090 32GB | more than 8 | — | — |
| L40S 48GB | 8 | 45.9 GiB | about 114,539 tokens |
| A100 80GB | 8 | 276.3 GiB | about 689,758 tokens |
| H100 80GB | 8 | 276.3 GiB | about 689,758 tokens |
| RTX PRO 6000 96GB | 4 | 55.9 GiB | about 139,505 tokens |
| H200 141GB | 4 | 217.9 GiB | about 543,956 tokens |
Serving
A vLLM launch line for the shape above. Check the model card for a runtime the publisher recommends.
vllm serve XiaomiMiMo/MiMo-V2.6-Pro-RL \ --tensor-parallel-size 8 \ --max-model-len 110592 \ --gpu-memory-utilization 0.90
Questions
The weights take about 280 GiB at the released precision. No single card in the table above holds that, so it needs several GPUs or a lower precision.
The repository declares mit. Read the licence text before commercial use: the name alone does not tell you what is allowed.
The configuration allows 1,048,576 tokens. Whether the whole window is usable depends on the memory left for the cache, which the hardware table works out per card.
This page is a reference, and Aiqre does not serve this model today. We deploy open-weight models on dedicated EU hardware on request, so if you want an OpenAI-compatible endpoint for it, email [email protected] with the model name and roughly what volume you expect.
Also in the library
MiMo-V2.6-Flash-RL is a 159.4 billion-parameter sparse MoE causal model for multimodal long-context reinforcement-learning tasks.
Nex-N2.5-Pro is a 396.8 billion-parameter MoE language model designed for long-horizon, agentic tasks with multimodal and compute…
GigaChat 3.5 Reasoning is a 438 billion-parameter mixture-of-experts language model trained with online RL for improved reasoning…
K2-Horizon-375B-A23B is an open-weight sparse MiE language model from IFM with 379 B parameters, 23 B active per token and a 512 …
Hosting
Tell us the model and roughly what volume you expect. We reply with a price and an OpenAI-compatible endpoint, with your prompts never stored.