Skip to content
NLEN
Illustration: a locked crate holding a model network, next to a calendar, a cloud server and a key that has not been used yet.

Inkling 2 Small: open weights promised for 'next week', still closed five days later

By Ivo Donker · — text: Claude (AI, Anthropic), checked against the sources

Can you download the Inkling 2 Small weights yet? No, not as of November 10, 2026. Thinking Machines Lab announced the model on November 5 and wrote that it intends to release the full weights "next week" under an Apache 2.0 license, after external safety testing and red-teaming. Five days later the same sentence is still on the page, with no date. Until then, Inkling 2 Small can only be used through Tinker, the company's training and sampling platform.

We read Thinking Machines Lab's announcement again on November 10. The topic also featured in the trending briefing of November 10.

What Inkling 2 Small is

Inkling 2 Small is the second generation of the Inkling models. It is a mixture-of-experts model with 594 billion parameters in total, 17 billion of them active per token. The model accepts images through a 292-million-parameter vision encoder and has adjustable effort. You can try it in the Tinker Playground and fine-tune it through Tinker; checkpoints only become downloadable once the weights are released.

The promise, verbatim

The page says: "We intend to make the full weights available under an Apache 2.0 license next week". The release depends on completing external safety testing and red-teaming. On November 10 there is no new date and no repository link. On the first results page of a search for "Inkling-2" on Hugging Face we found no Thinking Machines repository; only third-party repackagings of older Inkling models. Those are not an official release.

Benchmarks: the company's figures

Thinking Machines reports, among other things, a score of 31 on the Artificial Analysis Intelligence Index (the table says 30.6), 28.7 percent on HLE-Diamond and 81.6 percent on MMMU Pro. On SWEBench Pro v2 Hard (86.3 percent) and DeepSWE v1.1 (72.1 percent) the model ranks second in its own table. The company itself also names a limitation: on its own alignment evaluations Inkling 2 Small trails closed models such as GPT-6.1 Sol.

Why the date matters

An Apache 2.0 license would make Inkling 2 Small one of the most permissively usable large open models; we wrote earlier about license differences in the license debate around open models. But as long as the weights are not online, in practice it is a closed model on one platform. If you want it in your own stack, wait for the repository, not the announcement. The broader trend is in open-source LLM trends.

What we did not check

We did not test the model ourselves. The benchmarks are reported by Thinking Machines. On Hugging Face we only looked at the first results page of the search. As soon as the weights appear, we will update this article.

Checked on November 10, 2026.

This article was written with AI (Claude, Anthropic) and checked against the sources below.

Sources