Thinking Machines announces Inkling 2 Small, with open weights to follow

Thinking Machines Lab announced Inkling 2 Small, a mixture-of-experts model with 594 billion total and 17 billion active parameters that accepts image input and has variable reasoning effort. It is available now on the company’s Tinker platform, for sampling and fine-tuning, with closed weights. The company said it intends to release the full weights under an Apache 2.0 licence “next week”, after external safety testing and red-teaming, and that the model would reach further platforms “in the coming weeks”. The safety section describes the release as “pending further safety evaluations”.

This follows the sequence Thinking Machines set out in August: access through a hosted interface and fine-tuning API first, full weights later. It is a successor to Inkling-Small, which was released as an open-weight model.

The post gives an Artificial Analysis Intelligence Index score of 31 in its introduction, though its benchmark table lists 30.6 for the index’s v5.0, so the two figures differ slightly. Many other results in the comparison tables are the company’s own, alongside externally reported scores it attributes to Artificial Analysis, Scale AI, Vals and others.

On safety, Thinking Machines reported testing with an internal helpful-only variant and with more external testers than for earlier releases. It said Inkling 2 Small scored lower than Inkling and Inkling-Small on FORTRESS, a harmful-assistance evaluation, but trailed closed-weights models such as GPT-6.1 Sol on its alignment evaluations. These are the company’s own results.

Whether the weights ship as promised, and with what findings from the outside testers, is the open question. Cognition had announced the previous day a coding model built on the earlier Inkling-Small, and said it would not publish its weights for cyber reasons.