Overview
thinking-machines-lab multimodal open-weights MoE model: 975B total / ~41B active parameters, up to 1M-token context, Apache 2.0. Largest widely reported U.S. open-weights model as of Jul 2026; trails proprietary Claude/GPT and some Chinese open models on overall quality; weak factual accuracy (high hallucination rate on AA Omniscience).
Recent Developments
- 2026-07-15: Public release — Hugging Face weights (BF16 + NVFP4), tinker API/fine-tuning (2026-07-16-thinking-machines-inkling-official)
- Artificial Analysis Intelligence Index ~41 — top U.S. open-weights per THE DECODER (2026-07-16-thinking-machines-inkling-decoder)
- inkling-small preview (276B/12B)
- Pretrained on 45T multimodal tokens; day-0 support in transformers / SGLang / vLLM / llama.cpp
"Largest U.S. open-weights" and benchmark ranks are secondary/vendor-linked claims — cite Artificial Analysis / primary blog.
Related
- thinking-machines-lab
- mira-murati
- inkling-small
- tinker
- open-weight-models
- mixture-of-experts
- apache-2.0
- nemotron-3-ultra
- moonshot-ai
- us-open-weights-race-2026
- thinking-machines-inkling-975b-open-weights