Thinking Machines released Inkling 2 Mini on 20 November, an open-weight AI model that matches much larger rivals on coding tests and runs on an ordinary laptop with 32GB of memory.
The release matters because it breaks a link the AI industry has treated as fixed: serious coding models needed a data centre. Inkling 2 Mini does its heaviest work on a single machine, and its scores put it level with a 27-billion-parameter rival that costs about six times as much per task to run.
Key Points
- Thinking Machines released Inkling 2 Mini on 20 November 2026, with open weights under the Apache 2.0 licence.
- The model has 26.4 billion parameters in total and switches on about 3.4 billion for each token. Its middle layers run three times per token, giving 48 layer visits.
- It scored 48.2% on DeepSWE v1.1 and 79.2% on SWE-bench Verified, the best results among open models under 100 billion parameters on agentic coding tests, according to the company.
- Artificial Analysis scored it 20 on its Intelligence Index, tying Qwen3.8 27B, at $0.59 per task against $3.53.
- The checkpoint is about 20GB and runs on a Mac with 32GB of memory or an Nvidia DGX Spark, through Ollama, vLLM, SGLang and llama.cpp.
What happened
Thinking Machines Lab put the weights for Inkling 2 Mini on Hugging Face on 20 November. Anyone can download them, and the model is also available to try in the company’s Tinker Playground. The release is a looped transformer, a design that reuses the same layers instead of stacking more of them.
The model has 24 physical layers. For each token, the middle 12 layers run three times in a row, so a token passes through 48 layers in total. That gives Inkling 2 Mini the same depth as the far larger Inkling 2 Small while using about one twenty-second of its parameters. Thinking Machines says repeating the middle layers beat repeating the final layers on 27 of 28 tests it ran.
On the company’s own benchmarks, the model leads open models under 100 billion parameters on agentic coding. It resolved 48.2% of tasks on DeepSWE v1.1 and 79.2% on SWE-bench Verified, with 77.9% on the hard set of SWE-bench Pro v2. It also scored 46.5% on the Vals Index v2.2, 79.0% on BrowseComp, 85.3% on DeepSearchQA and 82.7% on MCP Atlas. The figures come from Thinking Machines and have not been repeated by outside labs yet.
The company showed the model working offline on a MacBook with an M5 Pro chip, where it built a playable flight simulator in a few minutes. In a second demo, one local copy of the model controlled 20 villagers in a simulated pixel town.
Building the model took a different route from the usual scale-up. Thinking Machines started with 384 experts in each layer and cut that to the 128 most used, then fed the trimmed model one trillion tokens distilled from a larger internal model so the surviving experts could take over the missing work. Fine-tuning and reinforcement learning followed, covering 2.2 trillion tokens over 2,000 steps. Training ran in NVFP4 precision on Nvidia GB300 hardware.
Why it matters
The first reason is practical. A model that runs on a laptop keeps data on the machine, works without an internet connection and needs no monthly API bill. For small teams and solo developers, that changes what they can build and where they can build it. Artificial Analysis measured the gap in money and time: Inkling 2 Mini cost $0.59 and took 9 minutes per Intelligence Index task, against $3.53 and 38.2 minutes for Qwen3.8 27B. Researchers at Stanford measured 0.703 intelligence-per-joule on an Nvidia DGX Spark, against 0.430 for Qwen3.6-35B and 0.159 for Gemma-4-26B.
The second reason is the trade-off it exposes. Looping layers adds computing power without adding stored knowledge, so factual recall stays at the level of other models in the 26-billion-parameter class. The model also uses about 38% more output tokens than Qwen3.8 27B for the same Intelligence Index score, even though it uses about 2.8 times less computing power overall. Its weaker results sit in everyday chat and strict instruction following, at 50.0% on MultiChallenge and 33.4% on ComplexConstraints, and larger models still lead on hard reasoning tests. Scores also depend heavily on the test setup, so developers should repeat the tests in their own setup before choosing a model.
The release lands in a week when the AI trade itself is under strain. Anthropic shares fell sharply on Friday as a selloff in AI stocks wiped more than $4 trillion off the Nasdaq. Interest in models that run locally is also growing as governments look harder at the data centres behind AI, and Washington is weighing a tax on AI data centres built abroad. A model that needs no data centre at all solves a different problem from the one Wall Street is pricing, but it is the problem many builders care about.
There is a caution attached to the design. Because the extra work happens inside the loops, the model can reason longer without writing more words down. That leaves less of its thinking visible in the text it produces, which makes it harder to check how it reached an answer. Thinking Machines released a research version with one to five adjustable passes so researchers can study that effect, and gave the interpretability lab Goodfire early access to the model’s inner workings.
What happens next
Developers will now test the model on their own machines, and the first independent results will show whether the company’s coding scores hold up outside its own test setup. vLLM shipped support on launch day, and the model already runs through Ollama, SGLang, Unsloth and llama.cpp.
The research version will get the closest attention. With one pass, the model worked on a problem for 16 seconds without writing anything down; with three passes, that stretched to 37 seconds before levelling off. For questions it answered correctly at every depth, its written reasoning fell from 3,703 tokens with one pass to 1,661 with five. That pattern is the core of the idea: more thinking, fewer words.
It was a week of bold AI claims. A separate claim that an AI system had proved the Riemann hypothesis is also awaiting scrutiny, and the same test applies to both: outside labs, on their own hardware, with their own tests. Inkling 2 Mini has cleared the first bar by shipping weights anyone can download. The next bar is replication.
Sources: Thinking Machines, Hugging Face, The Decoder, Artificial Analysis, AlphaSignal, Futurum Group. Published 22 November 2026.



