Skip to content

Thinking Machines ships Inkling 2 Mini, a 26B laptop model

A looped transformer that activates 3.4 billion of its 26.4 billion parameters per token, fits in about 20GB of memory, and leads every open-weight model under 100B parameters on two coding benchmarks.

Thinking Machines released Inkling 2 Mini on 20 November, a scaled-down, laptop-friendly sibling to its earlier Inkling 2 Small. It is a looped transformer: 24 layers arranged as four six-layer blocks, where the two middle blocks each run three times per token. That gives it 26.4 billion total parameters but only 3.4 billion active per token, rising to the equivalent of 5.4 billion once the looping is counted, so it reasons like a deeper model without the memory cost of one.

The whole thing fits in roughly 20GB, which Thinking Machines says is enough to run on a Mac with 32GB of RAM through Ollama, or on Nvidia’s DGX Spark. Weights are open and already up on Hugging Face, with support in Tinker, vLLM, SGLang and llama.cpp. On the company’s own numbers it scores 48.2% on DeepSWE v1.1 and 77.9% on SWE-bench Pro v2 Hard, and Thinking Machines states plainly that it “leads every open-weights model with less than 100B total parameters” on benchmarks including those two.

Why it matters: a coding-capable agentic model that runs on a 32GB Mac, rather than a GPU cluster, changes who can self-host instead of calling an API. The looped-layer design is also a reusable idea: other labs can borrow the trick of running a few layers repeatedly to add reasoning depth without inflating the parameter count you have to load into memory.

Share

Related reading

More from The Wire

All briefs