Weights stay closed until next week

The Thinking Machines Lab post Introducing Inkling 2 Small sets datePublished to 2026-11-05T00:00:00+00:00, which is 08:00 in Beijing on Nov. 5, exactly on this edition's backfill line and before the window that starts at 16:00, so this is backfill. The visible date is Nov. 5, 2026. dateModified is 2026-11-05T15:42:12-08:00, which is 07:42 in Beijing on Nov. 6. That is a later edit and does not pull the publication into the morning window. The discovery index says 2026-11-05T07:00:00Z, which is 15:00 in Beijing on Nov. 5, a different clock from datePublished. This column uses the page's own publication time.

The post says Inkling 2 Small is a mixture-of-experts model, with 594B parameters in total, 17B active, image inputs, and variable effort. On the Artificial Analysis Intelligence Index, the post gives a score of 31 and says it surpasses the lab's own previous models. The previous score is not in that sentence, so the size of the gap has no number. Training is summarized as more efficient pretraining and one large unified reinforcement-learning run. The procedure itself is not expanded here.

The weights are closed for now and are offered first through Tinker, including a trial on Tinker Playground and fine-tuning and sampling on Tinker. The post says it intends to release the full weights under an Apache 2.0 license next week, after external safety testing and red-teaming are finished. Next week is an intention, not a download that exists today. Checkpoints become downloadable when the weights are released. The safety section says the weight release is still waiting on further safety evaluation, split into two questions: how the model behaves for a general user, and how well it resists a malicious party removing safeguards. The extracted passage does not turn those two questions into a finished table of results.

The capabilities section calls long-horizon autonomous work the difference from the first generation, and says vision and agent behavior can use visual feedback to revise websites, games, and interactive three-dimensional designs. Named demonstrations include a city simulator built with Three.js and a from-scratch personal-computer emulator in Python that boots a real Linux kernel. Those are demonstrations the lab posted, not a scored board. The safety section frames the work as two questions: how the model behaves for a general user, and how well it resists a malicious party removing safeguards. The weight release is still pending further safety evaluations by the lab and its partners. On chemical, biological, radiological, and nuclear risk, the post says the knowledge and capability scores sit below the open-weight frontier. That sentence gives no score and does not say how the evaluation was run. This column does not add the method.

What is available now is closed weights on Tinker. The 31 is the lab's figure on the Artificial Analysis index, not an independent replication attached at the end. The image is the post's cover, a field of dots and loops, not that score table.