Sunday, 22 November 2026, story 2 of 7
Thinking Machines' Inkling 2 Mini loops its layers under Apache 2.0
The open model loops 12 of its 24 layers three times per token, and a GitHub project squeezes it to 5.38 GB for Apple silicon.

What happened
Thinking Machines has released Inkling 2 Mini, an open looped transformer under Apache 2.0, The Decoder reports. It runs 12 of its 24 layers three times per token, which the company says lets it reason longer without writing more text. The company claims the lead among open models under 100 billion parameters on agentic coding, with 48.2 percent on DeepSWE v1.1. Artificial Analysis scores it 20 on its Intelligence Index, tying Qwen3.8 27B at roughly six times lower cost and four times the speed.
Weights are on Hugging Face, with a variable-loop research version. Local runs need a Mac with 32 GB or more, or an Nvidia DGX Spark.
What it means
The benchmark claims are the company's, and the cost and speed comparison rests on one index, so test the model on your own tasks before swapping out Qwen3.8 27B. The practical news is the memory range. A GitHub project, LookAtMyQuant, lists four Apple silicon builds from 5.38 GB to 12.76 GB. In its own tests, the largest matched the full-precision model on 200 GSM8K and 164 HumanEval problems, while the smallest fell to 88.0 percent and 70.7 percent. Those are small samples, run by the author with reasoning off, so the HumanEval drop is the figure to weigh if you build coding agents. These builds target Apple silicon only. Independent runs of the quantized files, and of the variable-loop version, are the next thing to watch.
What to do: On a Mac, try the 8.9 GB or 12.8 GB build on your own tasks; the 5.4 GB build scored far lower on HumanEval.
Sources
Earlier on this: Ollama adds Thinking Machines' Inkling 2 Mini for Apple silicon Macs (21 November 2026).
Filed under Open source. Published .
"What it means" is analysis, not reporting. Found an error? Write to daily@vibitely.com.