DownloadMiMo-V2.5-Pro is Xiaomi’s flagship Mixture-of-Experts (MoE) model built for the most demanding agentic, software engineering, and long-horizon reasoning tasks. It features approximately 1.02 trillion total parameters with 42B activated per inference, balancing extreme capability with efficient execution. The model supports a 1 million token context window, enabling it to maintain coherence across long workflows involving thousands of tool calls and multi-step reasoning chains. Architecturally, it uses a hybrid attention system combining Sliding Window Attention and Global Attention to significantly reduce memory usage while preserving long-context performance. It also integrates multi-token prediction modules that accelerate inference and improve reinforcement learning efficiency. Trained on around 27 trillion tokens with FP8 mixed precision and refined through supervised fine-tuning, large-scale agentic reinforcement learning, and distillation.
Features
- 1.02T-parameter Mixture-of-Experts architecture
- 42B active parameters for efficient inference
- 1M-token context window for long-horizon workflows
- Hybrid attention (SWA + Global) for memory efficiency
- Multi-token prediction modules for faster generation
- Strong agentic performance with thousands of tool calls
- Trained on ~27T tokens with FP8 mixed precision
- Optimized for autonomous workflows and complex engineering tasks
Project Samples
