PublishersXiaomi MiMo

Xiaomi MiMo

10 documents read from this publisher.

XiaomiMiMo/MiMo-V2.6-Pro-RL · Hugging Face

first party release · model card · 2026-09-21 · 20 techniques

XiaomiMiMo/MiMo-V2.6-Flash-RL · Hugging Face

first party release · model card · 2026-09-21 · 16 techniques

Mixture of Experts Multi-Token Prediction Speculative decoding EAGLE DFlash Prefix-Conditioned On-Policy Distillation Groupwise Advantage Redistribution Groupwise Agentic Grading Groupwise Reward Synthesis Multi-Prefix Multi-Teacher On-Policy Distillation No Shared Experts Adversarial screening Asynchronous Group Relative Policy Optimization Environment hardening Self-correction cold start Verifier cross-checking

XiaomiMiMo/MiMo-V2.5-Pro · Hugging Face

first party release · model card · 2026-04-27 · 13 techniques

Mixture of Experts Multi-Token Prediction Hybrid Attention Multi-Teacher On-Policy Distillation Supervised fine-tuning Sliding Window Attention EAGLE Learnable attention sink bias Agentic reinforcement learning DeepEP FP8 mixed-precision training Global attention Multi-layer EAGLE

XiaomiMiMo/MiMo-V2.5-Base · Hugging Face

first party release · model card · 2026-04-27 · 19 techniques

Mixture of Experts Multi-Token Prediction Hybrid Attention Speculative decoding Multi-Teacher On-Policy Distillation Supervised fine-tuning FP8 Sliding Window Attention Chunked prefill Hybrid Sliding Window Attention Learnable attention sink bias Agentic reinforcement learning Attention data parallelism DeepEP FP8 mixed-precision training Global attention FlashAttention 3 Audio encoder initialized from MiMo-Audio Round-robin load balancing

XiaomiMiMo/MiMo-V2.5 · Hugging Face

first party release · model card · 2026-04-27 · 17 techniques

Mixture of Experts Multi-Token Prediction Hybrid Attention Speculative decoding Multi-Teacher On-Policy Distillation Supervised fine-tuning FP8 Sliding Window Attention Chunked prefill Hybrid Sliding Window Attention Learnable attention sink bias Agentic reinforcement learning DeepEP FP8 mixed-precision training Global attention FlashAttention 3 Vision Transformer (ViT)

Xiaomi MiMo-V2.5

first party release · official blog · 2026-04-22 · 11 techniques

Mixture of Experts Multi-Teacher On-Policy Distillation Supervised fine-tuning Reinforcement Learning Hybrid Sliding Window Attention Progressive context extension Agentic post-training multimodal pre-training Projector warmup Text pre-training Visual and audio encoders with lightweight projectors

Xiaomi MiMo-V2.6 Series: 3 New Models Officially Released

first party release · vendor docs · 2026-01-01 · 4 techniques

1M-token context window Multi-agent collaboration Joint multimodal input and understanding Optimization on real agent-harness environments

Xiaomi MiMo-V2.6 Series: 3 New Models Officially Released

first party release · vendor docs · 2026-01-01 · 5 techniques

Code-Driven Animation Iterative render refinement toward a playable result Literature review, hypothesis formation, simulation, and candidate shortlisting for materials research Multi-agent task decomposition and coordination Visual checking of generated renders

MiMo V2 6 technical report

technical report · 107 techniques

Hybrid Attention DFlash Asynchronous reinforcement learning Partial rollout Prefix-Conditioned On-Policy Distillation Agentic reinforcement learning Groupwise Advantage Redistribution Groupwise Agentic Grading Groupwise Reward Synthesis FP4 Quantization-Aware Training Freezing the MoE Router During Reinforcement Learning Multi-harness training Multi-Prefix Multi-Teacher On-Policy Distillation Rollout Routing Replay Autonomous Student Rollouts Decoupled control plane and data plane Deficit-corrected scheduling Predictive Rollout Dispatch Adaptive Rollout Concurrency Adaptive Rollout Scheduling Advantage Shaping Agent Loop Agent-centric mid-training Agent-centric multimodal data mixture Agent-centric rollout execution Alternating Row-Major and Column-Major Token Serialization Asynchronous cache offloading and restoration Asynchronous group-wise grading Atomic binary rubric evaluation Behavior rubrics Behavioral regularization Composite early-stop strategy Container-level network isolation CPU-resident optimizer states Cross-entropy objective Data Scheduler Data-parallel-first multimodal encoding Dialogue prefix matching Distribution-matched draft-model fine-tuning Domain-Specific GRPO Training

Install the latest SGlang from main branch

first party release · code repo · 12 techniques

Speculative decoding Multi-Token Prediction EAGLE Multi-Token Prediction Seamless Rollout Engine Cold-started SFT model Continuous rollout with asynchronous reward computation and early termination Data re-sampling strategy Multi-dimensional data filtering Rule-based verifier Test-difficulty-driven code reward Three-stage data mixture strategy