Thinking Machines Lab has released Inkling 2 Mini, an open-weights model small enough to run on a laptop. It has 26.4 billion total parameters and a memory footprint of about 20GB, and it uses a looped design to get more reasoning out of fewer tokens.
Among open-weights models under 100 billion parameters, the lab says this one ranks first on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%). Weights are out now, and the lab’s Tinker platform supports inference and customization.
How the looped design works
The architecture is Mixture-of-Experts, with 3.4 billion parameters active per token. Counting the repeated passes, that comes to 5.4 billion weight activations per token.
Its 24 layers form four blocks of six. Tokens pass through the first and last blocks once, and through the two middle blocks three times. The thinking is that abstract reasoning sits in the middle, with outer layers handling translation to and from internal representations. Testing backed that up: repeating the middle 12 layers outscored repeating the final 12 in 27 of 28 evaluations.
Layers also read earlier blocks’ outputs and learn how to blend them, a technique called attention residuals. A learned gate begins closed, so early on each loop acts like a model without looping, and it opens gradually. That kept training stable.
To shrink the model, the team removed two thirds of the experts used in its mainline recipe. The bulk of the weights are trained and served in NVFP4, apart from the last four layers in MXFP8.
What it does well
Long-horizon agentic work is where it does best. Agentic coding tests such as DeepSWE v1.1 and SWE-bench Pro v2 put it ahead of similar models, with only Qwen3.8-27B beating it elsewhere. On PMPP-Hard, a kernel-optimization test, it approaches the much larger Inkling 2 Small.
It also leads similar-size open models on BrowseComp, DeepSearchQA and MCP Atlas, which reward search and tool use. That matters because looping doesn’t add knowledge. Facts live mostly in total parameter count, and the model’s factuality score on AA-Omniscience matches other models of its size. Its answer is to look things up.
Stanford groups (the Hazy Research Lab and the Scaling Intelligence Lab) got early access and measured intelligence per joule on an NVIDIA DGX Spark. Inkling 2 Mini scored 0.703. Qwen3.6-35B-A3B scored 0.430, Gemma-4-26B-A4B scored 0.159 and Qwen3.5-27B scored 0.036. On a flight, the lab’s own researchers also ran it offline on a MacBook Pro, and it produced a working flight simulator within minutes.
Where it falls short
The lab is open about the trade-offs of shrinking the model and adding loops.
It isn’t the best small model for multi-turn conversation or heavily constrained instructions: MultiChallenge gives 50.0% and ComplexConstraints v1.1 gives 33.4%, and Qwen3.8-27B wins both. Bigger models also lead on reasoning-heavy tests like CritPt v2 and HLE-Diamond.
Knowledge-heavy benchmarks such as SimpleQA Verified, SimpleVQA and WorldVQA didn’t improve with extra passes, while long-context tests like GraphWalks BFS and MRCR v2 did.
Looping and oversight
Looping lets a model reason longer without writing that reasoning out, which could make its chain of thought less useful for monitoring. The lab checked this with a research-only variable-loop variant, released alongside the main model.
With no chain of thought, that variant’s time horizon grew from 16 seconds at one pass to 37 seconds at three. A serial reasoning benchmark kept improving up to ten loops, double the training count. Chains of thought shrank as loops increased yet stayed readable, and scores on agentic sabotage tests did not rise.
What to try next
The weights are open, so outside researchers can test the lab’s claims on their own hardware and tasks.
The efficiency numbers point to local hardware as the natural fit. The lab demonstrated a single offline MacBook running Inkling 2 Mini to manage all 20 residents of a simulated village, and it positions the model for running several persistent agents locally.
Developers can pull the open weights now. Chat-heavy or tightly constrained workloads should check those weaker scores first. Coding and tool-using agents are where it looks strongest.
Source: Thinking Machines Lab