⚡ Uncle Cat AI Radar
ModelsOpen SourceAgentsSafety

Thinking Machines releases 26B open-weight looped model

Thinking Machines released Inkling 2 Mini, a 26.4B open-weight model that loops internal layers to improve agentic coding on local hardware.

What happened

Thinking Machines Lab released Inkling 2 Mini, a 26.4-billion-parameter mixture-of-experts model with 3.4 billion active parameters and an Apache 2.0 license. Unlike a conventional transformer, its middle layers are run three times for each token, allowing the model to perform more serial computation without producing a long visible chain of thought.

The company says the model is available with full weights on Hugging Face and can run locally through Ollama, llama.cpp, Unsloth and other inference stacks. Its approximate 20GB memory footprint puts it within reach of laptops and desktop systems with at least 32GB of memory. Thinking Machines also released a variable-loop research model so researchers can study how additional passes affect reasoning and oversight.

Why the numbers matter

On the lab’s reported evaluations, Inkling 2 Mini reached 48.2% on DeepSWE v1.1, 77.9% on SWE-bench Pro v2, 79.0% on BrowseComp and 85.3% on DeepSearchQA. Those comparisons are supplied by the developer and use different harnesses across models, so they should not be treated as a universal ranking. The important result is the efficiency profile: the model claims competitive agentic performance while using far fewer active parameters than similarly scoring dense models.

Looping also creates a safety question. In the company’s experiments, chain-of-thought became shorter as passes increased but remained readable at this scale. A weaker monitor missed between 1.0% and 5.0% of simulated side-task attacks across different loop counts, while tool-only monitoring missed substantially more behavior.

Why it matters

Inkling 2 Mini is significant because it couples an unusual reasoning architecture with genuinely accessible weights. If the results transfer beyond the lab’s benchmarks, developers may get stronger local coding and tool-use models without buying frontier-scale infrastructure. The unresolved issue is whether internal computation remains reliably inspectable as looping depth and model capability grow.

Sources