MiMo-V2.5 — Xiaomi's Open-Source Omni-Modal AI Model

Xiaomi has released MiMo-V2.5. It is a native omni-modal model that handles text, images, video, and audio within a single architecture, and it's designed for agentic tasks. Since it's under the MIT license, commercial deployment and fine-tuning are also possible.

• MiMo-V2.5: A Sparse MoE structure that activates only 15B of its 310B total parameters
• MiMo-V2.5-Pro: A 1.02T/42B-scale Pro version focused on long-term consistency and complex software engineering.
• Hybrid Attention (SWA + GA 5:1, window 128): Reduces KV-cache storage by about 6x while supporting up to 1M tokens of context.
• Dedicated encoders: Equipped with a 729M-parameter ViT vision encoder and a 261M-parameter audio encoder.
• MTP (Multi-Token Prediction): 3 layers that speed up speculative decoding inference and improve RL efficiency.
• Training: Trained on roughly 48T tokens with FP8 mixed precision, followed by SFT, large-scale agentic RL, and Multi-Teacher On-Policy Distillation (MOPD).
• Deployment: Officially supported by SGLang and vLLM, offered in two versions, Base (256K) and Full (1M).

In the ClawEval results introduced by VentureBeat, the Pro model achieved a 63.8% success rate, a leading level among open-source models. It produced comparable results using 40-60% fewer tokens than Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4, and it scored 1581 points on GDPVal-AA (Elo), surpassing Kimi K2.6 and GLM 5.1.