Sunday, 22 November 2026

Inkling 2 Mini open model runs offline on 32GB Macs

Thinking Machines has released Inkling 2 Mini, an open-weights model of about 20GB that runs offline on Macs with 32GB of memory.

Key points

  • Inkling 2 Mini has 26.4 billion parameters and needs about 20GB of memory.
  • Ollama lists it as running on a Mac with 32GB or more.
  • The 48.2% DeepSWE score is the company's own result.

An AI model that scores near the top of its size class on coding tests can now run offline on a laptop with 32GB of memory. Thinking Machines Lab has released Inkling 2 Mini, an open-weights model that needs about 20GB of memory, so a 32GB Mac can hold it without a cloud connection.

The model is free to download and sits under the Apache 2.0 license, which allows commercial use. For anyone in Pakistan with a recent MacBook or a desktop with a large graphics card, that means a coding assistant that works without a subscription or an internet link, though the download itself is large.

Thinking Machines says the model has 26.4 billion parameters in total, with 3.4 billion active at a time, and that it loops its middle layers three times for every token. The idea is to let the model do more thinking internally and arrive at an answer with fewer written-out tokens. Counting the repeated layers, the company puts the work per token at 5.4 billion parameters.

The company reports 48.2% on the DeepSWE v1.1 coding benchmark and 46.5% on Vals Index v2.2, and says that is ahead of every open-weights model under 100 billion parameters. On SWE-bench Verified it lists 79.2%. These are the company’s own numbers, run under its own settings, so independent tests may land differently.

It is not the best at everything. Chat and instruction-following are weaker spots. By the company’s account, Qwen3.8-27B beats it on both MultiChallenge (it gets 50.0%) and ComplexConstraints v1.1 (33.4%), and larger models do better on hard reasoning tests such as HLE-Diamond. Looping also does not add stored knowledge, so the model leans on search and tools for facts.

Most of the weights were trained and are shipped in 4-bit NVFP4, with the last four layers in 8-bit MXFP8, which is how the file stays near 20GB without a separate shrinking step. Ollama lists it as running on a Mac with 32GB or more of memory, or on an NVIDIA DGX Spark. As a demo, the company’s researchers ran it offline on a MacBook Pro during a flight and had it write a working flight simulator within minutes.

Weights are on Hugging Face and the model can be pulled through Ollama. The company also released a research variant with a variable number of loops, which it used to study whether looping makes the model’s reasoning harder to monitor. At this size it reports no meaningful drop in how readable the written reasoning is.

TPRO Team

The Tech Prolonged newsroom desk. Stories under this byline cover AI, the web and general tech, and are reviewed by our editors before publishing.