# Thinking Machines releases open-weight Inkling 2 Mini

> The 26.4-billion-parameter model takes image and text input and can run on some laptops. Its repeated layers aim to add reasoning without storing more weights.

Oossa · 2026-11-20 · https://oossa.com/en/thinking-machines-releases-open-weight-inkling-2-mini

Thinking Machines Lab released Inkling 2 Mini on November 20, 2026. It is an open-weight model that accepts text and images, supports a one-million-token context window, and is intended to be small enough for local use. Artificial Analysis says it is available under the Apache 2.0 license; Thinking Machines says it is available for inference and customization through its Tinker service.

The model’s unusual feature is that it runs 12 of its 24 layers three times for each token. The repeated layers reuse the same weights, while messages pass between their runs. That gives the model more computation per token without storing three separate copies of those layers. Thinking Machines says the design is meant to help with tasks that require several steps, such as coding agents.

## How big is it, and what can it do?

The parameter counts depend on how repeated work is counted. Thinking Machines lists 26.4 billion total parameters, 3.4 billion active parameters, and 5.4 billion weight activations per token when looping is included. Artificial Analysis counts 5.4 billion active parameters including the repeated passes, and says the model uses 3.1 billion unique weights per text token. Hugging Face describes it as roughly 26 billion total and 3 billion active parameters.

Artificial Analysis found Inkling 2 Mini matched Qwen3.8 27B at its xhigh setting and Nanbeige4.3-33B on its Intelligence Index, while using fewer active parameters. Its task-level estimate puts the model at $0.59 and nine minutes per Index task, compared with $3.53 and 38.2 minutes for Qwen3.8 27B. Those cost estimates use Thinking Machines’ published API prices; they are not estimates of the cost of running it locally.

Thinking Machines reports scores of 48.2% on DeepSWE v1.1 and 77.9% on the hard 51-task SWE-bench Pro v2 test. The company cautions that evaluation setups differ between models. Hugging Face also says further validation is needed, so these results should not be read as proof that the model is generally better than alternatives.

## Can it run on a personal computer?

Thinking Machines says the model’s memory footprint is about 20GB and describes it running offline on a MacBook Pro with an M5 Pro chip and 48GB of memory. Hugging Face lists support for Transformers, llama.cpp and MLX, among other tools. That gives technically inclined users routes to try it locally, but it does not mean it will run well on every laptop.

Thinking Machines also released a research version that can use a variable number of loops. The company says fewer loops make inference faster but may reduce answer quality; it used that version to study how looping affects performance and the ability to monitor a model’s reasoning. That experimental checkpoint is distinct from the standard Inkling 2 Mini release.

## The facts

- Inkling 2 Mini has 26.4 billion total parameters, according to Thinking Machines Lab.
- Its 12 middle layers run three times per token, reusing the same weights.
- Artificial Analysis reports a one-million-token context window and text-and-image input.
- Thinking Machines reports 48.2% on DeepSWE v1.1 and 77.9% on SWE-bench Pro v2 Hard.

## Why it matters

For developers with a sufficiently capable laptop, open weights and support for local inference tools offer a way to try a model without sending every prompt to a hosted chatbot. The company reports a roughly 20GB memory footprint, but performance on ordinary consumer laptops remains unclear.

## Sources & references

1. [Thinking Machines has released Inkling 2 Mini, an open weights looped transformer](https://artificialanalysis.ai/articles/inkling-2-mini) – Artificial Analysis, 2026-11-20
2. [Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup](https://huggingface.co/blog/inkling2mini) – Hugging Face, 2026-11-20
3. [Introducing Inkling 2 Mini](https://thinkingmachines.ai/news/inkling-2-mini/) – Thinking Machines Lab, 2026-11-20

Last updated: 2026-11-20
