huggingface.co web signal

Thinking Machines ships variable-loop Inkling research model

TL;DR

  • Thinking Machines Lab released a 26.4B-parameter research model on November 20, 2026 whose middle 12 layers can be looped 1 to 5 times at inference.
  • Each extra pass adds about 1.0B effective active parameters per token, taking per-token active weights from 3.4B at one pass to 7.4B at five.
  • The model was post-trained at three passes and the card warns it may behave unpredictably at other counts, so it is explicitly not for production.

Thinking Machines Lab put a research-only sibling of its Inkling 2 Mini on Hugging Face on November 20, 2026, with the loop count exposed as an inference-time knob.

The architecture description, from the model card: "A 24-layer decoder-only transformer with sparse Mixture-of-Experts (MoE) feed-forward layers. The middle 12 layers run a variable number of times with shared weights, chosen at inference time; the model was trained with 1 to 5 passes." Total weights come in at 26.4B with 3.4B active, and "each extra pass through the looped layers adds about 1.0B effective active parameters per token" — 3.4B at one pass, 5.4B at three, 7.4B at five.

The framing the lab puts on the release is specific: "We are releasing it with open weights to support research on looped models. Because the number of loops can be varied on a single set of weights, the model allows cleaner comparisons of how additional looped computation affects a model's internal computation and overall capabilities."

The training schedule behind that: "For the rest of midtraining, it was trained with 1 to 5 passes through the looped layers, sampled in a 1:2:6:2:1 ratio, and it was then post-trained with 3 passes." The post-training is why the card calls the weights a "research artifact" and warns the model "may behave unpredictably at pass counts other than 3." A companion Hugging Face blog post is blunter about the maturity: "We tested this variant, and it works, but we haven't evaluated it in depth."

Two researchers we track flagged the drop within hours of it going live.

Shared on Bluesky by 2 AI experts