Thinking Machines Lab, the startup founded by former OpenAI chief technology officer Mira Murati, released a new open-weights model on November 20 called Inkling 2 Mini — and the interesting part is not the size, but what the model does with it. Inkling 2 Mini is a 26.4-billion-parameter mixture-of-experts transformer, but instead of passing each token through every layer exactly once, it runs its middle 12 layers three times per token, reusing the same weights on each pass.
The arithmetic is unusual. Only 3.4 billion parameters activate in a single pass; once looping is counted, each token touches about 5.4 billion weight activations. The 24-layer network is split into four blocks of six, with layers 7 through 18 executed three times and the first and last six run once, producing 48 layer applications from 24 stored layers. A learned gate merges the output of each loop pass into the next, and the company reports that looping the middle layers beat looping the final layers on 27 of 28 evaluations.
The practical payoff is footprint. Most weights were trained and are deployed natively in NVFP4, with the final four layers kept in MXFP8, which keeps the whole model at roughly 20GB — small enough for a Mac with 32GB of unified memory or an NVIDIA DGX Spark. The context window reaches 1 million tokens, with text and image input. Thinking Machines demonstrated the model running fully offline on a MacBook Pro during a flight, where it built a playable flight simulator in minutes.
On benchmarks, the company says Inkling 2 Mini leads every open-weights model under 100 billion total parameters on DeepSWE v1.1 (48.2 percent) and Vals Index v2.2 (46.5 percent), both agentic-coding evaluations — figures produced by the vendor itself on a mini-swe-agent v2.4.6 harness and not yet independently reproduced. On the Artificial Analysis Intelligence Index it scores 20.1, tying Qwen3.8-27B, and Artificial Analysis's own cost measurements put Mini at $0.59 and 9.0 minutes per index task versus $3.53 and 38.2 minutes for Qwen3.8-27B. The trade-offs are candid: the model trails Qwen3.8-27B on multi-turn conversation (MultiChallenge 50.0 percent) and constraint-heavy instructions (ComplexConstraints v1.1 33.4 percent), and its knowledge scores sit at the level expected for its size.
Alongside the fixed model, the lab released a variable-loop research variant trained with one to five passes, and its findings will interest safety researchers more than benchmark chasers. On that checkpoint, the time horizon for tasks completed without any chain of thought rose from 16 seconds at one pass to 37 seconds at three, and serial-reasoning accuracy kept improving through ten passes — twice the training depth. Meanwhile the visible chain of thought grew shorter as loops were added, though the company reports it remained legible, and the model did not score higher on agentic-sabotage tests at greater loop depth. If models start reasoning internally through loops rather than writing their reasoning down, the monitorability of chain-of-thought — one of the few practical oversight tools available today — becomes an open question on real weights rather than a theoretical one.
Training details released with the model show the loop was a deliberate bet: the team began with 384 experts per layer, pruned to 128 at a mid-training checkpoint, distilled about 1 trillion tokens from a larger internal Inkling model, and finished with supervised fine-tuning and reinforcement learning totalling 2.2 trillion tokens. The company estimates a non-looping model would have been 7.5 percent cheaper to train but needed roughly 140 percent more inference compute to match the looped version's loss on long agentic traces.
Weights are on Hugging Face under Apache 2.0, with day-zero serving support from vLLM, SGLang, Unsloth, NVIDIA Dynamo and TokenSpeed, plus Ollama access that remains a pre-release build. OpenRouter lists the model at $0.15 per million input tokens and $0.50 per million output. For the industry, the release matters less for its leaderboard placement than for what it normalizes: a frontier-adjacent lab openly shipping an architecture where compute is spent in depth rather than breadth, and inviting the safety community to study what happens to oversight when the reasoning stops being written down.
Comments (0)
Log in to join the discussion
Log InNo comments yet