SourcesThinking Machines Lab
publisher · 5 sources
Thinking Machines Lab
5 sources read from this publisher, newest first, with the methods each one describes.
thinkingmachines/Inkling-NVFP4 · Hugging Face
first party release · model card · 2026-07-15 · 14 methods
Mixture of expertsHybrid Sliding-Window and Global AttentionTop-k expert routingNative multimodal architectureDiscrete audio token encodingHierarchical patch encoderJoint multimodal decoding in a shared hidden spaceMulti-turn open-ended external red-teamingRefusal-suppressed capability evaluationData cleaning, processing, and modificationDeduplication and filteringbenchmarking against public frontier modelsdefense-in-depth with layered input/output classifiersPre-release safety evaluation
Inkling: Our Open-Weights Model
first party release · official blog · 2026-07-15 · 19 methods
Mixture of expertsHybrid Sliding-Window and Global AttentionAsynchronous reinforcement learningContinuously controllable reasoning effortEncoder-free multimodal architecturePython tool for visual tasksLearned relative attentionSFT bootstrapping with synthetic dataAbstention-aware factual QA rewardHybrid Muon and Adam optimizationPost-training on diverse domainsRandomized tool set and schema during trainingRL for calibration with proper scoring rulesRL with rubric and claims gradersSafety training to an internal specificationShort convolutions in attention and residual branchesSigmoid router with auxiliary-loss-free load balancingTraining to resist censorshipWeight decay coupled to learning rate squared
Inkling Model Card
first party release · model card · 2026-07-15 · 10 methods
Hybrid Sliding-Window and Global AttentionNVFP4Shared expertDiscrete audio token encodingHierarchical patch encoderMulti-turn open-ended external red-teamingRefusal-suppressed capability evaluationData cleaning, processing, and modificationDefense-in-depth for safetyInput/output classification with moderation tools
thinkingmachines/Inkling · Hugging Face
first party release · model card · 2026-07-15 · 15 methods
Mixture of expertsHybrid Sliding-Window and Global AttentionShared expertTop-k expert routingDiscrete audio token encodingHierarchical patch encoderJoint multimodal decoding in a shared hidden spaceMulti-turn open-ended external red-teamingRefusal-suppressed capability evaluationSynthetic data generationDeduplication and filteringApplication-layer safeguards (content filtering, rate limiting, monitoring)Dangerous-capability uplift assessmentHuman oversight and review for high-stakes model outputsInput/output moderation classifiers layered around the model (defense-in-depth)
Latest stable release from PyPI
software documentation · code repo · 6 methods