MiMo V2.6

Xiaomi’s MiMo-V2.6 family — MIT-licensed native-omnimodal MoE with 1M context. Flagship open checkpoint is MiMo-V2.6-Pro-RL (1.02T total / 42B activated); Flash-RL is 309B / 15B activated. Hosted API ids: mimo-v2.6-pro, mimo-v2.6-flash, mimo-v2.6-pro-ultraspeed. Xiaomi publishes an Artificial Analysis Intelligence Index score of 46 for Pro; that figure is vendor-published and is not a DataAIHub ranking.

Vision · Audio · Tool calling · Thinking · Coding

Last reviewed: 22 September 2026

Official pricing →

Overview

Capabilities

Vision: Yes
Audio: Yes
Tool calling: Yes
Thinking: Yes
MCP: No
Coding: Yes
Structured output: Yes

Technical specifications

Provider: Xiaomi
License: MIT
Context window: 1.0M
Parameters: Pro MoE 1.02T/42B active; Flash MoE 309B/15B active
Architecture: Sparse MoE (SWA + global attention; MTP drafter)
Release: 2026-09
Modalities: Text, Image, Video, Audio
Vision: Yes
Audio: Yes
Tool calling: Yes
Thinking: Yes
MCP: No
Open weights: Yes
API: Yes
Pricing (input): Xiaomi MiMo Open Platform (verify live rates)
Pricing (output): Xiaomi MiMo Open Platform (verify live rates)

Supported modalities

Text · Image · Video · Audio

Context window

1.0M (1,048,576 tokens)

Pricing

Input: Xiaomi MiMo Open Platform (verify live rates)

Output: Xiaomi MiMo Open Platform (verify live rates)

Xiaomi says V2.6 keeps V2.5 API list prices and adds Pro UltraSpeed. Self-host via MIT weights. Do not treat third-party rate tables as official.

Availability

API: Yes

Chat UI: Yes

Open weights: Yes

Open weights: XiaomiMiMo/MiMo-V2.6-Pro-RL and XiaomiMiMo/MiMo-V2.6-Flash-RL (MIT; collection XiaomiMiMo/mimo-v26). Xiaomi also released MiMo-V2.6-Distill-Qwen-9B plus RL task environments. Hosted: MiMo Open Platform, AI Studio, MiMo Code, MiMo Desktop, and listed on OpenRouter. Model-card serving recipes: SGLang (preferred) and a vLLM image based on the MiMo-V2.5 recipe — the Flash card says stable vLLM may lag. Do not assume Ollama/LM Studio day-one support.

Use cases

Open-weight omnimodal agents (text, image, video, audio)
Long-context coding and tool-using agents
Self-hosted MIT-licensed deployments
Hosted MiMo API when not self-hosting

Strengths

MIT license on Pro-RL and Flash-RL weights
Native text/image/video/audio with 1M context
Documented SGLang and vLLM serve recipes on the model cards

Limitations

Pro self-host footprint is multi-node class on the published SGLang recipe
vLLM support may lag; card points at a MiMo-V2.5 image
Newer ecosystem than Llama/Qwen; Xiaomi is not yet a DataAIHub company page
Vendor benchmark/AA claims are not DataAIHub scores

Related guides

Related benchmarks

Related research

MiMo-V2.6: Scaling Up Reinforcement Learning for Self-Improvement
Original Paper
MiMo-V2.6-Pro-RL (Hugging Face)
Original Paper
MiMo-V2.6-Flash-RL (Hugging Face)
Original Paper