Key takeaways
- Thinking Machines Lab released Inkling 2 Mini on Nov. 20 with open weights, and it is available on the company's Tinker platform.
- The model has 26.4 billion total parameters, 3.4 billion active per token, and loops its two middle blocks of layers three times for each token.
- The company reported a 48.2% score on DeepSWE v1.1 but said the model trails Qwen3.8-27B on multi-turn chat and complex instructions.
In this story 5 sections
Thinking Machines Lab released Inkling 2 Mini on Nov. 20, an open-weights AI model that the startup says fits on a consumer laptop and tops every other open model under 100 billion total parameters on tests including the DeepSWE v1.1 coding benchmark. The weights are public, and the model can be used today for inference and customization on Tinker, the company's own platform, according to the Thinking Machines announcement.
The company described it as a model "small enough to run on your laptop." Its total memory footprint is about 20GB.
Why Inkling 2 Mini loops its middle layers
The unusual part is the architecture. Inkling 2 Mini is a looped transformer with 24 layers arranged as four blocks of six. The first and last blocks run once, but the two blocks in the middle repeat three times on each token. Thinking Machines says the repetition gives the model more room to reason without writing out extra tokens, an advantage it says shows up most in agentic coding.
The lab said it tested where to put the loop. Repeating the middle 12 layers beat repeating the last 12 on 27 of 28 evaluations it ran, the company said.
To shrink the model, Thinking Machines pruned two thirds of the experts from its mainline recipe. Most weights were trained and are served natively in the NVFP4 format, with the final four layers in MXFP8.
Inkling 2 Mini by the numbers
All scores below are figures the company reported, and it notes that evaluation setups differ from model to model.
- Size: 26.4 billion total parameters, 3.4 billion active per token, and 5.4 billion weight activations per token once the loops are counted.
- DeepSWE v1.1: 48.2%, against 42.2% for Qwen3.8-27B from Alibaba Group (NYSE: BABA), the closest rival in the company's table.
- SWE-bench Verified (bash only): 79.2%.
- SWE-bench Pro v2 Hard: 77.9%.
- Vals Index v2.2: 46.5%.
- BrowseComp: 79.0%.
Two figures we worked out from those numbers: only about 12.9% of the model's parameters are active for any given token (3.4 billion divided by 26.4 billion), and its DeepSWE lead over Qwen3.8-27B is 6 percentage points, or roughly 14% in relative terms (48.2 divided by 42.2).
What running it locally looks like
Thinking Machines said its researchers used the model on a flight, with an Apple (NASDAQ: AAPL) MacBook Pro (M5 Pro, 48GB) in airplane mode, and had it write a working flight simulator game in minutes without an internet connection.
Outside testers measured efficiency too. Researchers at Stanford's Hazy Research Lab and Scaling Intelligence Lab, who had early access, scored Inkling 2 Mini at 0.703 on an intelligence per joule measure while running it on DGX Spark hardware from Nvidia (NASDAQ: NVDA). The company listed 0.430 for Qwen3.6-35B-A3B and 0.159 for Gemma 4 26B A4B, an open model from Google, part of Alphabet (NASDAQ: GOOGL), on the same measure.
Where Thinking Machines says it falls short
The company was direct about weak spots. For long back and forth conversations and prompts that stack many rules, Qwen3.8-27B does better: Inkling 2 Mini scored 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, trailing that model on both. It also lags bigger models on hard reasoning tests such as CritPt v2 and HLE-Diamond.
Looping does not let the weights hold more facts, Thinking Machines said, because stored knowledge depends mostly on total parameter count. The company says the model makes up for that by leaning on search and tools.
A research version for studying oversight
Alongside the main release, the lab published a research variant whose number of loops can be changed. Because looped models can reason internally without writing their steps down, the company used it to test whether people can still follow what the model is doing through its chain of thought.
With one pass, the variant's time horizon on tasks without written reasoning was 16 seconds; with three passes it reached 37 seconds, about 2.3 times longer by our math. Thinking Machines said the written reasoning got shorter as loops were added but remained readable, and that scores on its agentic sabotage tests did not rise.
For background on how models like this work, see our explainer on what a large language model is. Another recent open release is covered in our report on Cloudflare's Clairic models, and more coverage is on our AI news page.
Share this story
Spotted an error? Tell us


