Thinking Machines releases Inkling 2 Mini, an open model for laptops
The company says the model scores 48.2% on DeepSWE v1.1, a test for code tasks, more than all other open models below 100 billion parameters.
Thinking Machines Lab, with Mira Murati as its head, released Inkling 2 Mini on 20 November 2026. It is a model with open weights, and a person can download its files. The company says the model is fast, good at agent work (software that does tasks with tools) and sufficiently small to operate on a laptop, at about 20GB. It reports 48.2% on DeepSWE v1.1.
How it works
The model operates its middle layers three times for each token (a small unit of text). This gives more reasoning in short written steps. Only 3.4 billion of its parameters, the values in the model, are active.
The numbers
On DeepSWE v1.1 the company reports the top score of small open models. Artificial Analysis, an external tester, finds a task costs $0.59, against $3.53 for Qwen3.8-27B. These are task figures, not laptop tests.
What is not known
The model trails Qwen3.8-27B in long chats and in tasks with many rules. Looping lets a model do more work without written steps, and this is not easy to monitor. The company says the written steps stayed short and easy to read, and sabotage test scores did not increase.
Insight: An open small model follows a closed large model
Inkling 2 Small launched on 5 November with closed weights, until external safety tests end. Some users found it too large. Mini has open weights from day one, and it also operates on Tinker, the platform of the lab. This reads as a move to win developers with a small open model while the large model stays on the platform of the lab. Monitor the release of the Small weights.
Sources
Posted