SourcesMistral AI
publisher · 4 sources
Mistral AI
4 sources read from this publisher, newest first, with the methods each one describes.
Mistral Large 4 (Le Chonk) is now open weight | Mistral
first party release · official blog · 2026-10-29 · 19 methods
Mixture of expertsThinking modeReinforcement learningFine-tuningNative multimodal architectureTensorRT-LLMBrowser-based testing and iterative repairCustomization using proprietary organizational dataCustomization with proprietary cybersecurity data and policiesCybersecurity guardrailsCybersecurity-specialized model variantExpert-customer collaboration to build and deploy tailored AI systemsManaged-service deployment with private Azure deployment optionsModel customizationReinforcement-learning pipelineScientific-tool answer cross-checkingSelf-hosted model deployment with inspectable, customizable weightsSFT checkpoint initialization for RLTask-specific model distillation or optimization into smaller variants
mistralai/Mistral-Large-4-1T-A52B · Hugging Face
first party release · model card · 2026-10-29 · 14 methods
Mixture of expertsThinking modeNVFP4DSparkShared expertTemperature and nucleus sampling configurationThree-token speculative draft depthGranular Mixture-of-Experts with routed and shared expertsMinimal tool-set designPixtral-style variable-resolution vision encoderRecurrent Markov draft headStrict JSON Schema outputSystem-prompt guidance for tool useThree-layer dense MLA speculative drafter
Introducing Mistral Large 4 | Mistral
first party release · official blog · 2026-10-06 · 10 methods
Supervised fine-tuningReinforcement learningAsynchronous reinforcement learningMultimodal inputAutoscaling rollout actorsComposable multi-method reward verificationHybrid instruction-and-reasoning MoELow-staleness optimization for long-trajectory RLOff-policy drift reductionShared composable interface for RL tasks
technical report
technical report · 115 methods
Mixture of expertsThinking modeMulti-Teacher On-Policy DistillationAsynchronous reinforcement learningDeepSeek Sparse AttentionFP8Sliding Window AttentionMuonDSparkYaRNAdamWBatch-size warmupAdversarial-agent reward-hacking auditFull-vocabulary logit distillationAdaptive Rollout ConcurrencyAgent-based synthetic-data generation pipelineContext parallelismDropping zero-advantage rollout groupsGroup-relative advantage estimationHeterogeneous training mixtureKernel fusionLeRoPEMixed-precision NVFP4 routed-expert quantizationRouter replayShared environment implementation across RL, evaluation, and synthetic data generationTaxonomy-guided generation of simulated workplace environmentsAbs-max quantizationActivation CPU offloadingActor-pool orchestration for asynchronous RLAppend-only episode recordsAsynchronous execution of supporting training workloadsAutomated health checks and crash analysis for training operationsAutoscaling supporting servicesBackpressure- and concurrency-controlled rollout admissionBlock verificationBlockwise scales in generator and trainerCentral schedulerCentralized trajectory bin-packingCISPO-style REINFORCE objectiveCommon agent-environment interaction loop