Thinking Machines Lab, the startup Mira Murati founded after leaving OpenAI, has released Inkling 2 Mini, an open-weights model small enough to run offline on a laptop. Its trick is architectural: the middle of the network runs three times for every token, so a model with 26.4 billion parameters reasons like a bigger one without needing a data center. On the lab’s own coding benchmarks it now sits ahead of Qwen3.8-27B, the model most local setups use as the yardstick.
A 24-layer model that behaves like a 48-layer one
Inkling 2 Mini is a mixture-of-experts model with 24 layers. The first six and the last six run once; the twelve in between run three passes with shared weights, which gives 48 layer visits per token. Later passes do not simply take the previous output as input. Each one starts again from the output of the first block and receives gated “messages” from the pass before, a design the Hugging Face launch post walks through in detail.
Per token the model activates 3.4 billion parameters, or 5.4 billion once the repeated passes are counted. The weights ship in the precision they were trained in: NVFP4 for most layers and MXFP8 for the last four, about 20 GB in total. The license is Apache 2.0, and support for transformers, llama.cpp, MLX, vLLM and SGLang landed on day one.
The lab also reports a training trade-off worth noting. In its comparison, a non-looping version cost 7.5% less compute to train but needed 140% more inference compute to reach the same loss on long agentic traces. Looping, in other words, shifts cost from every future query to a one-time training bill, which is exactly the trade a local model wants.
Our comparison: more coding score per active parameter
We put the lab’s published scores next to Qwen3.8-27B, a dense model that uses all 27 billion parameters on every token:
- DeepSWE v1.1: 48.2% for Inkling 2 Mini against 42.2% for Qwen3.8-27B, a 6-point lead with about one fifth of the active compute per token (5.4B versus 27B).
- SWE-bench Verified (bash only): 79.2% against 77.8%.
- Terminal Bench v2.1: 71.6% against 73.0%, one of the few coding rows Qwen still wins.
- Energy: in an early test by Stanford researchers on an NVIDIA DGX Spark, Inkling 2 Mini scored 0.703 on their intelligence-per-joule metric, about 1.6 times the 0.430 of Qwen3.6-35B-A3B.
All of these numbers come from Thinking Machines and its testing partners. Evaluation setups differ between models, so treat the gaps as indicative until independent leaderboards rerun them.
Second looped model in a week
Inkling 2 Mini is not the only model betting on repeated layers. Three days ago SenseTime released SenseNova-Loop-9B, which runs a block of 24 layers twice and fits into about 4.9 GB at 4-bit. The two designs point in different directions: SenseTime optimised for the smallest possible footprint and, by its own figures, still trailed Qwen3.8-27B on repository-level coding, while Thinking Machines spent four times the memory and claims to pull ahead. Looping is turning from a research curiosity into a real option for local models, and the next few weeks of independent testing should show which balance holds up.
On the other end of the open-weights spectrum, Tencent this week opened its 770B Hy4 flagship, a model no laptop will ever run. Inkling 2 Mini competes for a different job: the coding agent that lives on your own machine.
Who should try it, and what to watch
For developers who already run local coding agents on a Mac with 48 GB of unified memory or a desktop AI box, Inkling 2 Mini is worth a test against their current Qwen setup, especially for agentic tasks that call tools or search. The vLLM day-0 notes add a practical caveat: every looped layer keeps its own cache per pass, so the model holds 48 KV caches for 24 layers. Long contexts therefore eat more memory than the parameter count suggests, and a 20 GB footprint does not mean a 32 GB laptop can use the full 1M-token window.
The lab is also open about weak spots. Inkling 2 Mini scores 50.0% on MultiChallenge and 33.4% on ComplexConstraints, behind Qwen3.8-27B on both, so long chats with many rules are not its strength. Looping adds reasoning, not memory for facts: its non-hallucination rate on AA-Omniscience is 63.4% against Qwen’s 71.4%, which is why the lab pushes it toward search and tool use. Output outside English is weaker too, according to the model card.
What is still open: independent benchmark runs, real-world speed on consumer laptops at long context, and how the looped design behaves when the community fine-tunes it. For a wider view of how the wrapper around a coding model shifts results, see our look at harness versus model tests.



