Thinking Machines’ Laptop Model Scores 48.2% on DeepSWE

Thinking Machines Lab released Inkling 2 Mini on Friday, an open-weights model it says is “small enough to run on your laptop”, and reported a 48.2% score on the DeepSWE v1.1 coding benchmark, per its announcement.

The company says that leads every open-weights model under 100 billion parameters on that test. The weights are on Hugging Face under an Apache 2.0 licence.

Business Pill · Free Cash Flow

An evergreen Business Engineer explainer. It teaches the general idea only and does not discuss this model.

The key insight: Thinking Machines says Inkling 2 Mini, a 26.4B-parameter open-weights looped transformer of about 20GB, scores 48.2% on DeepSWE v1.1, 6.0 points ahead of the next model in its own table (our arithmetic).

A Model That Runs Its Middle Layers Three Times

Inkling 2 Mini is a looped transformer. Its model card says the middle 12 of its 24 layers “run three times with shared weights”, giving 48 layer applications per token.

It has 26.4 billion parameters, 3.4 billion of them active per token, or 12.9% (our arithmetic). Counting the loops, the company puts it at 5.4 billion weight activations per token.

The checkpoint is about 20GB. The card lists a single Nvidia DGX Spark or Apple silicon with 32GB or more of memory as supported hardware.

DeepSWE v1.1 scores as reported by Thinking Machines; evaluation setups vary by model.
DeepSWE v1.1 scores as reported by Thinking Machines; evaluation setups vary by model.

The Benchmark Claim

In the company’s own table, Inkling 2 Mini scores 48.2% on DeepSWE v1.1, against 42.2% for Qwen3.8-27B: a 6.0-point lead (our arithmetic). Thinking Machines notes that “evaluation setups vary by model.”

On SWE-bench Verified it reports 79.2%, against 77.8% for Qwen3.8-27B.

What Looping Does Not Buy

The company says looping “doesn’t enable its weights to store more facts”, so knowledge still tracks total parameter count. It leans on search and tools to close that gap.

A research variant with a variable number of loops saw its no-chain-of-thought time horizon rise from 16 seconds at one pass to 37 seconds at three, the company says.

The Structural Read

As we read it, looping trades extra compute per token for fewer stored parameters, which is what lets a laptop-sized model compete on coding.

As we read it, the company’s own caveat on facts means tool use, not size, carries its knowledge work.

Thinking Machines Lab, 20 Nov 2026

“While looping improves the model’s reasoning, it doesn’t enable its weights to store more facts”

Three Implications

FOR DEVELOPERS the weights are on Hugging Face under Apache 2.0, about 20GB (model card).

FOR HARDWARE MAKERS the card lists a DGX Spark or a 32GB Apple silicon machine as supported.

FOR BENCHMARKS the scores are company-reported; independent results are not out yet.

What Is Not Established

All benchmark scores here are Thinking Machines’ own; we have not seen independent results yet.

Where It Fits

It follows the first Inkling in July. Our recent Thinking Machines coverage includes its Nvidia FP4 results and its hiring.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The AI Supercycle

The Business Engineer map of who builds, finances and captures value in AI.

Read the Map of AI →

The Bottom Line

Thinking Machines says its 26.4-billion-parameter, roughly 20GB Inkling 2 Mini scores 48.2% on DeepSWE v1.1, 6.0 points ahead of the next model in its own table.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Every figure is from Thinking Machines’ announcement and its Hugging Face model card. The benchmark scores are the company’s own. The 6.0-point lead and the 12.9% active share are our arithmetic.

Sources: Thinking Machines Lab, ‘Introducing Inkling 2 Mini’, 20 Nov 2026 · Inkling 2 Mini model card, Hugging Face · FourWeekMBA, Jul 2026: the first Inkling

Scroll to Top
FourWeekMBA