Verified, then logged.
Sun 22 Nov 2026 · 00:23 KST / 15:23 UTC한국어 | ENSearch /

02 / IT · Analysis

Thinking Machines Lab releases Inkling 2 Mini weights

In Thinking Machines Lab’s tests, a research model scored higher on long-context tasks when its middle layers ran more times.

Thinking Machines Lab released about 20GB of Inkling 2 Mini weights on Hugging Face. The model card gives that estimate for the download and lists runtime memory requirements separately.

Thinking Machines Lab demonstrated Inkling 2 Mini on an offline laptop. The model reuses the same weights each time its middle layers run.

The company tested different numbers of passes using a research version of the model. More passes improved scores on long-context tasks.

Middle layers repeat for each token

According to Thinking Machines Lab, the middle layers run three times per token.

The company said the first and last blocks each run once. On repeat passes, the middle blocks also process the output from the previous pass.

The model card lists support for Apple silicon devices with at least 32GB of memory.

Scores improve on long-context tasks

In the company’s tests, extra passes improved the research model’s long-context scores. Scores on knowledge tasks did not improve.

The company reported that the model’s written reasoning grew shorter as it made more passes. It remained readable throughout the tests.

Thinking Machines Lab said Inkling 2 Mini used search and other tools to fill gaps in its knowledge.