Thinking Machines Lab (Jul 15, 2026) released Inkling, its first open-weights foundation model.
Key facts (primary announcement):
- Architecture: Mixture-of-Experts transformer, 975B total parameters, 41B active
- Context: up to 1M tokens
- Pretraining: 45 trillion tokens of text, images, audio and video
- Multimodal: natively processes text, images, audio; video in pretraining mix
- Optimization: hybrid Muon (large matrix weights) + AdamW; parameter schedules inspired by prior work
- Family: Inkling-Small preview (276B total / 12B active MoE) with different latency trade-off; often matches or exceeds larger sibling on many benchmarks due to improved data/recipe
- Availability: Tinker platform (fine-tuning/API) with 64K and 256K context options; 50% limited-time discount; full weights on Hugging Face (original + NVFP4 for NVIDIA Blackwell)
- Positioning: first in a family; intended for customization/fine-tuning rather than only API rental
Cross-check: Apache 2.0 licensing and Hugging Face distribution corroborated by Hugging Face blog and secondary coverage (Axios, THE DECODER, The Register).