Thinking Machines Launches Inkling 2 Mini
Original: Inkling 2 Mini
Why This Matters
Looped transformer architecture directly links compute depth to oversight quality at small scale.
Thinking Machines Lab released Inkling 2 Mini on Nov 20, 2026—a 26.4B-parameter looped transformer MoE model scoring 48.2% on DeepSWE v1.1, topping all open-weights models under 100B parameters.
Inkling 2 Mini is a Mixture-of-Experts model with 26.4B total parameters, 3.4B active, and 5.4B weight activations per token when accounting for its looping design. The key architectural twist: the model reruns its middle transformer layers three times per token, letting it reason longer without generating visible chain-of-thought tokens. That matters for both performance and oversight. On benchmarks, it leads all open-weights models under 100B parameters—48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2—while running faster than comparably capable models. Most weights are trained and deployed natively in NVFP4 low-bit quantization, with the final four layers handled separately. Thinking Machines also released a variable-loop research variant alongside the main model to study monitorability. That variant's no-CoT time horizon grew from 16 seconds at one loop to 37 seconds at three, and performance on serial reasoning benchmarks kept climbing through ten loops—double its training depth. Crucially, chain-of-thought outputs got shorter with more loops but stayed readable, and the model showed no uptick on agentic sabotage tests as loop count increased. The model is available on Hugging Face and can run on a laptop.