PublishersThinking Machines Lab

Thinking Machines Lab

5 documents read from this publisher.

thinkingmachines/Inkling-NVFP4 · Hugging Face

first party release · model card · 2026-07-15 · 14 techniques

Mixture of Experts Hybrid Attention Discrete token encoding for audio Hierarchical patch encoder Joint multimodal decoding in a shared hidden space Multi-turn open-ended external red-teaming Refusal-suppressed variants for capability estimation Token-level expert routing Data curation Training data deduplication and filtering benchmarking against public frontier models defense-in-depth with layered input/output classifiers Native multimodal encoding Pre-release safety evaluation

Inkling: Our Open-Weights Model

first party release · official blog · 2026-07-15 · 19 techniques

Mixture of Experts Hybrid Attention Encoder-free multimodal architecture Python tool Relative attention SFT bootstrapping with synthetic data Abstention-aware reward for factual QA Controllable thinking effort via system message and per-token cost Hybrid optimization with Muon and Adam Large-scale asynchronous RL Post-training on diverse domains Reinforcement Learning for Calibration with Proper Scoring Rules RL with Rubric and Claims Graders Safety training to an internal specification Short convolutions in attention and residual branches Sigmoid-based MoE router with auxiliary-loss-free load balancing Tool-set and schema randomization Training to resist censorship Weight decay coupled to learning-rate squared

Inkling Model Card

first party release · model card · 2026-07-15 · 10 techniques

Hybrid Attention NVFP4 Discrete token encoding for audio Hierarchical patch encoder Multi-turn open-ended external red-teaming Refusal-suppressed variants for capability estimation Data curation Shared Experts Active on Every Token Defense-in-depth for safety Input/output classification with moderation tools

thinkingmachines/Inkling · Hugging Face

first party release · model card · 2026-07-15 · 15 techniques

Mixture of Experts Hybrid Attention Discrete token encoding for audio Hierarchical patch encoder Joint multimodal decoding in a shared hidden space Multi-turn open-ended external red-teaming Refusal-suppressed variants for capability estimation Shared Experts Active on Every Token Training data deduplication and filtering Application-layer safeguards (content filtering, rate limiting, monitoring)Dangerous-capability uplift assessment Human oversight and review for high-stakes model outputs Input/output moderation classifiers layered around the model (defense-in-depth)Synthetic data generation and augmentation Top-6 expert routing

Latest stable release from PyPI

software documentation · code repo · 6 techniques

LoRA fine-tuning Benchmark evaluation framework Claude Code skills for Tinker Documenting evaluation configuration Inline training evaluators Pre-commit hooks with ruff