> ## Content Index
> Fetch the complete content index at: https://venture-block.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Thinking Machines releases Inkling 2 Mini, a looped open model that runs on a laptop
- URL: https://venture-block.com/thinking-machines-inkling-2-mini-looped-open-model/
- Published: 2026-11-21T07:02:52.000Z
- Updated: 2026-11-21T07:02:52.000Z
- Description: Mira Murati's lab released an Apache-licensed, 20GB looped mixture-of-experts model that runs its middle layers three times per token and, by the company's count, leads open models under 100B parameters on agentic coding.
- Author: Venture Block
- Tags: AI, Tech

Thinking Machines Lab, the AI startup led by former OpenAI chief technology officer Mira Murati, on Nov. 20 released [Inkling 2 Mini](https://thinkingmachines.ai/news/inkling-2-mini/?ref=venture-block.com), an open-weight model built to run on a laptop. The model's checkpoint takes up about 20GB, and the company says it leads every open-weight model under 100 billion total parameters on benchmarks including the agentic coding test DeepSWE v1.1, where it resolved 48.2% of tasks.

The weights are on [Hugging Face under an Apache 2.0 license](https://huggingface.co/thinkingmachines/Inkling-2-Mini?ref=venture-block.com), which allows commercial use and fine-tuning. The model is also available for inference and customization on Tinker, the company's fine-tuning platform.

## A model that reads its own middle layers three times

The technical news is the architecture. Inkling 2 Mini is a looped transformer: of its 24 layers, the middle 12 run three times for every token, reusing the same weights on each pass. According to the [model card](https://huggingface.co/thinkingmachines/Inkling-2-Mini?ref=venture-block.com), that gives the model 48 layer applications per token while storing only 24 layers' worth of weights.

It is also a mixture-of-experts model, with 26.4 billion total parameters and 3.4 billion active per token. Each token is routed to 8 of 128 experts in every layer. Thinking Machines counts the extra passes separately: each additional trip through the looped layers adds about 1 billion effective active parameters, so the model does work equivalent to roughly 5.4 billion active parameters per token.

The passes do not simply feed into one another. As the [vLLM team explained in its day-zero support post](https://vllm.ai/blog/2026-11-20-inkling2-mini?ref=venture-block.com), all three passes start from the output of the first six layers, and each later pass reads gated "messages" from the pass before it. Each pass keeps its own key-value cache, so the 24-layer model maintains 48 KV caches. Hugging Face's engineers, in their [launch write-up](https://huggingface.co/blog/inkling2mini?ref=venture-block.com), said that in their own informal tests the model was at least comparable with Qwen 3.8 27B, while noting more validation is needed.

## Why the size matters

Most of the weights were trained and are served natively in NVFP4, a 4-bit floating-point format, with the last four layers in MXFP8\. That is what keeps the footprint near 20GB. The model card lists supported local hardware as a single Nvidia DGX Spark, Apple silicon machines with 32GB or more of memory, or other GPUs with at least 32GB of VRAM. Thinking Machines said its researchers ran it offline on a MacBook Pro during a flight and had it build a playable flight simulator in a few minutes.

Context length is up to 1 million tokens, and the model accepts images as well as text. Support was available at launch across common open-source runtimes, including vLLM, SGLang, Ollama, llama.cpp and Unsloth, according to the company.

On the company's published benchmark table, Inkling 2 Mini scored 79.2% on SWE-bench Verified and 77.9% on the hard subset of SWE-bench Pro v2, ahead of Google's Gemma 4 31B and Alibaba's Qwen3.8-27B on both. It is not ahead everywhere: Qwen3.8-27B scored higher on Terminal Bench v2.1, GDPVal-AA and AutomationBench-AA, and the two tied at 20.1 on the Artificial Analysis Index. All of these are figures reported by Thinking Machines, and the company notes that evaluation setups vary by model.

The release comes about two weeks after the lab's much larger [Inkling 2 Small](https://www.marktechpost.com/2026/11/05/thinking-machines-inkling-2-small-release/?ref=venture-block.com), a 594 billion-parameter mixture-of-experts model with 17 billion active parameters. It lands in a crowded month for model launches, after Google's [Gemini 4 Hydrogen](https://venture-block.com/google-gemini-4-hydrogen-replaces-flash/) and Anthropic's [Claude Fable and Mythos 5.2](https://venture-block.com/anthropic-claude-fable-mythos-5-2-launch/).

## The monitoring question

Looping lets a model do more reasoning internally without writing it out, which raises a concern for safety researchers who oversee models by reading their chain of thought. Thinking Machines released a second, research-only checkpoint that can run with a variable number of loops so outside researchers can study that trade-off.

In the company's tests on that variant, the model's no-chain-of-thought time horizon, a measure of how long a task it can finish without visible reasoning, rose from 16 seconds at one pass to 37 seconds at three, then roughly leveled off. The company said its chain of thought got shorter with more loops but stayed legible, and it did not see higher scores on tests of agentic sabotage. Its conclusion is limited to this scale: looping increased what the model can do silently, but did not meaningfully reduce how monitorable its reasoning was.

The model card is also direct about limits. Inkling 2 Mini is weaker than Inkling 2 Small on factual recall, can get stuck in long reasoning loops while debugging, and may delete or overwrite files without confirmation when used as an agent. Thinking Machines recommends running agentic workloads in a sandbox.

The lab said it is working with the National Deep Inference Fabric to make the looped models easier for interpretability researchers to access, and that the interpretability lab Goodfire had early access to them.