MiMo-V2.6-Pro-MOPD
huggingface.co- Category
- Developer Tools
- Rank
- No. 2395Tools index
Previous survey · No. 2307 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Models: Train & Run
- Date
About
MiMo-V2.6-Pro-MOPD is a multimodal large language model from Xiaomi's MiMo team that handles text generation, vision-language understanding, audio, video understanding, and agentic tool calling with long-context support, released under the MIT license. Its architecture combines an LLM backbone with a dedicated MiMo vision encoder, audio encoders, and a speculative decoder, and it can be served through vLLM or SGLang for OpenAI-compatible inference.
What it does
Xiaomi's model card describes an open-weight mixture-of-experts model, 1.02 trillion parameters with 42 billion active per token, that reads text, images, video and audio across a 1M-token context. This checkpoint adds a distillation pass from several specialist teacher models, which the card says curbs a habit of repeating the same tool call.
Stated on the product site
- Architecture
- Sparse mixture of experts, 1.02T total and 42B activated parameters
- Context length
- 1M tokens
- Input modalities
- Text, image, video and audio
- Lineage
- MOPD upgrade of the MiMo-V2.6-Pro-RL checkpoint
- Self-hosted serving
- SGLang and vLLM, with launch commands on the card
- Weight formats
- Safetensors in BF16, F32, F8_E4M3 and U8 tensor types
- Hosted access
- AI Studio, MiMo Code, MiMo Desktop, MiMo Open Platform API and OpenRouter
Not stated on the site
- The page text does not state a license, so the MIT license in the catalogue entry could not be confirmed from this read.
- The card refers to a figure comparing tool-call repetition before and after this checkpoint, but no numbers appear in the page text.
- No benchmark results appear on the page.
- The launch commands shard the model across many GPUs, but the page does not state minimum hardware or memory requirements.
Written from the product site at huggingface.co.
Tags
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.