SourcesHugging Face
publisher · 4 sources
Hugging Face
4 sources read from this publisher, newest first, with the methods each one describes.
Welcome Le Chonk: Mistral Large 4 on Hugging Face
technical report · third party analysis · 2026-10-29 · 22 methods
Mixture of expertsThinking modeDSparkShared expertLow-Rank AdaptationTop-k expert routingPrefix cachingSparse attentionReinforcement Learning with Verifiable RewardsMulti-head Latent AttentionAssistant-only lossCross-layer KV-cache sharingLeRoPESequence packingAlternating RoPE/NoPE layersDeepSeek Sparse Attention indexerFine-tuning on traces from Mistral Large 4Pixtral vision encoderReasoning cannot be disabledSpatial patch-merge projectorTensor and expert parallelismTRL
SmolVLA: Efficient Vision-Language-Action Model trained on Lerobot Community Data
first party release · official blog · 2025-06-03 · 28 methods
Asynchronous inference and executionInterleaved Self-Attention and Cross-AttentionCross-attentionPre-training on community-contributed robotics datasetsAction expertCausal self-attentionManual camera-viewpoint normalizationMultitask fine-tuningReduced action-expert hidden sizeSmall pretrained VLM backboneStandardized action-oriented task descriptions from frames and labelsTask-specific post-trainingVisual token reductionCap visual tokens at 64 per frameChunk fusionFlow matching objective for continuous action predictionGlobal-image-only runtime inputIntermediate VLM-layer featuresIntermediate-layer-limited action-expert conditioningManual review of candidate robotics datasetsNoise action sequences and predict a correction vectorPixelShuffle image compression for visual-token reductionPretraining on general manipulation dataRewrite task instructions under 30 characters beginning with an action verbRobotics dataset selection by frame count, visual quality, and task coverageSkip upper vision-language model layers for inference efficiencyThreshold-triggered inferenceVision-language model backbone
SmolVLA: A vision-language-action model for affordable and efficient robotics
research paper · technical report · 2025-06-02 · 42 methods
Asynchronous reinforcement learningFlow matchingAsynchronous inference and executionBF16 mixed-precision trainingBidirectional attentionInterleaved Self-Attention and Cross-AttentionCross-attentionFlow-matching action expertPre-training on community-contributed robotics datasetsAction-chunk executionCausal attentionCausal self-attentionImitation learningManual camera-viewpoint normalizationMultitask fine-tuningReduced action-expert hidden sizeSmall pretrained VLM backboneStandardized action-oriented task descriptions from frames and labelsVisual token reductionAction chunk sizes of 10–50 time stepsChunked action predictionCommunity robotics dataset filtering by metadata and qualityCross-attention-only action expertDropping unused camera views during trainingFixed-shape batching with excess-frame truncationFrame-and-instruction prompting for automatic task annotationFrequent observation updates during chunk executionFrozen-backbone action-expert fine-tuningJoint-space similarity filtering of observationsL1 regression loss on predicted action chunksMulti-task trainingOverlapping action-queue aggregationPass sensorimotor states directly to the action expertPer-action observation and prediction in simulationPixel-shuffle visual token reductionPyTorch JIT compilation with torch.compileQueue-threshold-triggered observation captureSensorimotor-state projection into VLM token spaceSkip every second VLM layerSynchronous action-chunk inference
lerobot/smolvla_base · Hugging Face
software documentation · model card · 2025-06-02 · 4 methods