SourcesXiaomi MiMo
publisher · 10 sources
Xiaomi MiMo
10 sources read from this publisher, newest first, with the methods each one describes.
XiaomiMiMo/MiMo-V2.6-Pro-RL · Hugging Face
first party release · model card · 2026-09-21 · 20 methods
Mixture of expertsMulti-Token PredictionHybrid Sliding-Window and Global AttentionvLLMPrefix-Conditioned On-Policy DistillationAsynchronous Group Relative Policy OptimizationAdversarial-agent reward-hacking auditGroupwise Advantage RedistributionGroupwise Agentic GradingGroupwise Reward SynthesisSGLangMoE without shared expertsMulti-Prefix Multi-Teacher On-Policy DistillationEnvironment hardening during reinforcement learningFive-layer SWA MTP drafterSelf-correction cold startVerifier cross-checkingGlobal attention with a dense feed-forward network in the first Transformer blockMixing tasks and harnesses in one RL batchSingle mixed-domain RL run
XiaomiMiMo/MiMo-V2.6-Flash-RL · Hugging Face
first party release · model card · 2026-09-21 · 16 methods
Mixture of expertsMulti-Token PredictionSpeculative decodingEAGLEDFlashPrefix-Conditioned On-Policy DistillationAsynchronous Group Relative Policy OptimizationAdversarial-agent reward-hacking auditGroupwise Advantage RedistributionGroupwise Agentic GradingGroupwise Reward SynthesisMoE without shared expertsMulti-Prefix Multi-Teacher On-Policy DistillationEnvironment hardening during reinforcement learningSelf-correction cold startVerifier cross-checking
XiaomiMiMo/MiMo-V2.5-Pro · Hugging Face
first party release · model card · 2026-04-27 · 13 methods
XiaomiMiMo/MiMo-V2.5-Base · Hugging Face
first party release · model card · 2026-04-27 · 19 methods
Mixture of expertsMulti-Token PredictionMulti-Teacher On-Policy DistillationSupervised fine-tuningHybrid Sliding-Window and Global AttentionFP8Sliding Window AttentionSpeculative decodingAgentic reinforcement learningAttention data parallelismChunked prefillGlobal attentionHybrid Sliding Window AttentionDeepEPFP8 mixed-precision trainingLearnable attention sink biasFlashAttention-3MiMo-Audio-initialized audio encoderRound-robin load balancing
XiaomiMiMo/MiMo-V2.5 · Hugging Face
first party release · model card · 2026-04-27 · 17 methods
Mixture of expertsMulti-Token PredictionMulti-Teacher On-Policy DistillationSupervised fine-tuningHybrid Sliding-Window and Global AttentionFP8Sliding Window AttentionSpeculative decodingAgentic reinforcement learningChunked prefillGlobal attentionHybrid Sliding Window AttentionDeepEPFP8 mixed-precision trainingLearnable attention sink biasFlashAttention-3Vision Transformer (ViT)
Xiaomi MiMo-V2.5
first party release · official blog · 2026-04-22 · 11 methods
Xiaomi MiMo-V2.6 Series: 3 New Models Officially Released
first party release · vendor docs · 2026-01-01 · 4 methods
Xiaomi MiMo-V2.6 Series: 3 New Models Officially Released
first party release · vendor docs · 2026-01-01 · 5 methods
MiMo V2 6 technical report
technical report · 108 methods
Hybrid Sliding-Window and Global AttentionAsynchronous reinforcement learningAgentic reinforcement learningDFlashPrefix-Conditioned On-Policy DistillationPartial rolloutAdversarial-agent reward-hacking auditGroupwise Advantage RedistributionGroupwise Agentic GradingGroupwise Reward SynthesisMXFP4 Quantization-Aware TrainingAutomated large-scale data synthesis and environment constructionFreezing the MoE router during reinforcement learningMulti-harness reinforcement learningMulti-Prefix Multi-Teacher On-Policy DistillationRollout Routing ReplayAdaptive Rollout ConcurrencyDecoupled control plane and data planeDeficit-corrected schedulingDropping zero-advantage rollout groupsDynamic sampling of informative rollout groupsFive-layer SWA MTP drafterIsolated resettable sandboxes for agent rolloutsPersistent multi-tenant harness poolPredictive Rollout DispatchAdaptive Rollout SchedulingAdvantage shapingAgent LoopAgent-centric mid-trainingAgent-centric multimodal data mixtureAgent-centric rollout executionAlternating Row-Major and Column-Major Token SerializationAsynchronous group-wise gradingAsynchronous KV-cache offloadingAtomic binary rubric evaluationAutonomous Student RolloutsBackground-thread offloading of blocking rollout workBehavior rubricsBehavioral regularizationComposite early-stop strategy that zeroes reward and masks affected contexts
Install the latest SGlang from main branch
first party release · code repo · 12 methods
MTP speculative decodingMulti-Token PredictionEAGLEMulti-Token PredictionSeamless Rollout EngineCold-started supervised fine-tuning modelContinuous rollout with asynchronous rewards and early terminationData re-sampling strategyMulti-dimensional data filteringRule-based verifiersTest-difficulty-driven code rewardThree-stage data mixture strategy