← In the News

Thinking Machines releases an open-weights coding model that runs on a 32GB Mac

Introducing Inkling 2 Mini · Thinking Machines Lab · November 20, 2026

Machine-readable Download Markdown

Inkling 2 Mini is a Mixture-of-Experts model with 26.4B total parameters, 3.4B of them active. Its memory footprint is about 20GB. The lab says it runs on any Mac with at least 32GB of RAM through Ollama, and the full weights are on Hugging Face. It is a looped transformer: the middle layers run three times for every token. On the lab's reported scores it leads comparable open models on agentic coding benchmarks, including 79.2% on SWE-bench Verified (bash only) and 48.2% on DeepSWE v1.1. The lab's own chart caption says "Evaluation setups vary by model," and we have not seen the scores reproduced independently.

The looping raises an oversight question. A looped model does more computation per token before it writes anything, so less of its reasoning may show up in its chain of thought. The lab tested this on LinuxArena, where an attacker model tries to finish a malicious side task during a realistic software engineering task without the monitor noticing. It used a helpful-only fine-tune, because the released model refuses many of the side tasks, and a weaker model as the monitor. Its finding: "With the monitor able to access only tool calls and outputs, it misses a significant number of the attacks." It saw no clear trend as loop passes increased, and describes the whole monitorability analysis as preliminary.

Why it matters: Running an open-weights coding model on a 32GB Mac, offline, is now a supported setup, which gives a factory a place for work that cannot leave the machine. How well it codes inside your own harness is something to measure rather than read off the table. The LinuxArena finding applies to any agent setup: what the monitor is allowed to see decides what it catches.