Inkling 2 Mini relies on loops to fit on a laptop

Thinking Machines Lab releases Inkling 2 Mini, a 26.4 billion parameter open-weight model designed for agentic use and local execution. The model is available on Hugging Face, Tinker, and several inference tools.

Thinking Machines Lab has released Inkling 2 Mini, an open-weight model designed to be fast, reason in agentic tasks, and run on a laptop. Available since November 20, 2026, it comes with a variable-loop research version and immediate deployment on Hugging Face and Tinker, with fine-tuning and chat in Tinker Playground. The goal is simple: reduce size without sacrificing key uses. Not always easy, even for a model that likes to go in circles.

Un modèle plus compact, mais pas rabougri

Inkling 2 Mini is a looped transformer, meaning a model that replays certain layers multiple times for the same token to extend its internal reasoning without generating more text. Of its 24 layers, the middle 12 run three times per token, while the first six and last six run only once. Thinking Machines reports a total of 26.4 billion parameters, including 3.4 billion active, with 5.4 billion weight activations per token once the loop is accounted for. The model is also described as light enough to run locally on a MacBook Pro, with a footprint of about 20 GB.

This positioning directly addresses criticisms aimed at Inkling 2 Small, considered too massive with its 594 billion parameters. Here, the logic is reversed: less volume, more efficiency, and a clear target for developers who want to test a powerful model without heavy infrastructure. The idea is not to beat the giants on all fronts, but to become practical again for everyday use.

Des performances ciblées sur le travail agentique

According to Thinking Machines, Inkling 2 Mini tops open-weight models under 100 billion parameters on several tests, including DeepSWE v1.1 with 48.2% and Vals Index v2.2 with 46.5%. On Artificial Analysis, it scores 20.1 points, tied with Qwen3.8-27B [to be verified according to sources], while being about four times faster and six times less costly per task on the Intelligence Index, according to figures published by the same source.

The model also stands out on BrowseComp, DeepSearchQA, and MCP Atlas, evaluations measuring tool use, research, and agentic tasks. This aligns with the lab’s strategy: compensate for a smaller model memory by better use of tools. However, Thinking Machines also clearly states that Inkling 2 Mini lags behind Qwen3.8-27B on multi-turn chat and very constrained instructions. The model therefore does not aim to do everything, which at least spares it from promising the moon in pocket format.

Une distribution large dès le lancement

The launch is not limited to a weight publication on Hugging Face. Thinking Machines announces support from day one on SGLang, vLLM, Ollama, Unsloth, NVIDIA Dynamo, and TokenSpeed for open-source inference, as well as RL fine-tuning support on SkyRL, Miles, and NVIDIA AutoModel. The model is also accessible in Tinker for inference and adjustment, with a chat interface in Tinker Playground.

This wide distribution is as important as the raw results. In today’s open model market, concrete access to the right tools often makes the difference between a noticed release and a merely commented one. Here, Mira Murati’s lab clearly aims to establish a working standard, not just publish another benchmark.

Un objet d’étude sur le raisonnement en boucle

Thinking Machines presents Inkling 2 Mini not only as a usable product. The company also publishes a variable-loop research variant to study what looped reasoning changes inside the model. According to its measurements, the horizon time without chain-of-thought goes from 16 seconds at one pass to 37 seconds at three passes, while the chain-of-thought becomes shorter but remains readable. Agentic sabotage tests showed no clear increase with the addition of loops.

The subject goes beyond pure performance. If a model can “think more and write less,” monitoring its reasoning becomes more delicate as it progresses. Thinking Machines half-admits this by sharing this research version with actors like the National Deep Inference Fabric and Goodfire, to closely study the behavior of looped models. For interpretability researchers, this is an interesting signal. For the rest of the market, it is a discreet reminder that AI sometimes loves to take detours to get to the right place faster.

By the numbers

  • 26.4 billion — total parameters of Inkling 2 Mini, according to Thinking Machines.
  • 24 layers — model architecture, including 12 layers repeated three times per token.
  • 20.1 points — score on the Artificial Analysis Intelligence Index, according to cited sources.
  • 1 million tokens — context window announced by secondary sources.
  • 20 GB — approximate memory footprint of the full model, according to Thinking Machines.

To read