By Paul Derrington, Jr.
CEO | Chief Engineer | Research Engineer | Editor, D-CSIL
3 min read
Key Takeaways
Thinking Machines released Inkling 2 Mini on November 20. The open-weight 26.4B model runs its middle layers three times per token. The company reports strong scores on selected coding benchmarks, and Artificial Analysis found a tie with Qwen3.8-27B (xhigh) on its Intelligence Index. Thinking Machines sells access through Tinker; its own tests were not independently verified.
The trick: run the middle layers three times
Inkling 2 Mini runs the middle 12 of its 24 layers three times per token. That means 48 layer applications using 24 stored layers. Thinking Machines reports the middle-loop design beat looping the last 12 layers on 27 of 28 evaluations in an internal comparison; that result was not independently verified.
Thinking Machines lists 26.4B parameters, 3.4B active parameters and 5.4B weight activations per token counting the extra passes. Most weights were trained and released in 4-bit NVFP4, with the final four layers in 8-bit MXFP8. This describes the native weights, not the quality of every downloadable conversion. The model accepts text and images.
What the benchmarks say
In its own launch tests, Thinking Machines reports 48.2% on DeepSWE v1.1 and 79.2% on SWE-bench Verified (bash only) for Inkling 2 Mini, compared with 42.2% and 77.8% for Qwen3.8-27B. These are company-reported results with varying evaluation setups and were not independently verified.
Artificial Analysis independently scored Inkling 2 Mini at 20.1 on its Intelligence Index, tied with Qwen3.8-27B (xhigh) and Nanbeige4.3-33B. For tasks in that benchmark, it reported $0.59 and 9.0 minutes for Inkling versus $3.53 and 38.2 minutes for Qwen, using provider pricing. Inkling used about 38% more output tokens than Qwen in that comparison. Artificial Analysis estimated about 36% of Qwen’s compute using an active-parameters-times-tokens proxy, not measured hardware compute.
Where it falls short
Thinking Machines reports 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1. Its launch prose says Qwen led on both, but its table puts Qwen at 30.1% on ComplexConstraints. The company’s own figures were not independently verified, and we do not treat the disputed comparison as settled.
The company also reports that written reasoning became shorter as it added passes to a separate variable-loop research model. It found no trend toward more agentic sabotage in the tests it ran. These are preliminary, company-run results, not independent findings or evidence that the released model universally writes less or is easier to monitor.
How you get it, and what it runs on
The official weights are on Hugging Face under Apache 2.0. Thinking Machines sells inference and fine-tuning through Tinker. Its model card lists Apple silicon Macs with at least 32 GB of memory among supported setups. That does not mean the full million-token model context fits in 32 GB; memory use depends on the model file, runtime, context and other workloads.
The company showed a local coding demo on a 48 GB M5 Pro MacBook Pro, where the model generated a playable flight simulator. This was a company demonstration, not an independently measured speed or a guarantee for other Macs.