Thinking Machines Lab releases Inkling-Small, a distilled open-weight model

Thinking Machines Lab released Inkling-Small, a smaller companion to Inkling, the open-weight model it had shipped two weeks earlier. Inkling-Small is a sparse mixture-of-experts model with 276 billion total parameters and 12 billion active per token — each token routed to 6 of 256 specialist experts plus 2 shared ones — packaged as a 42-layer transformer that accepts text, image and audio input and produces text output. It supports a 1-million-token context window and was released under the Apache 2.0 licence, with weights downloadable from Hugging Face in BF16 and NVFP4 formats.

The model card reported that Inkling-Small performs close to, and on some measures ahead of, the larger Inkling despite being under a third of its size: the full model scored higher on AIME 2026 (97.1% versus 95.5%), while Inkling-Small scored higher on HMMT (90.2% versus 86.3%). Thinking Machines said its safety evaluations found the smaller model’s risk profile consistent with existing open-weight systems, which the company said was the basis for releasing its weights in full rather than holding them back.

The release came alongside a staged-release framework the company published days later, arguing that a model should only reach full open release once evidence showed it did not extend the frontier of dangerous capability — a test it said both Inkling models had passed.