In brief
- The model runs the middle 12 of its 24 physical decoder layers three times.
- It accepts text and image inputs, with a context window of up to 1 million tokens.
- Deployment support is available for Transformers, llama.cpp, MLX, and SGLang.
- The Decoder reports an Apache 2.0 license and a local deployment option on Macs with at least 32 GB of memory.
Thinking Machines Lab introduced Inkling 2 Mini, a multimodal AI model, on November 20, 2026, in a Hugging Face blog post. Developed for local use, the model can process text and image inputs and offers a 1 million-token context window. According to Hugging Face’s announcement, the model has approximately 26 billion physical parameters and around 3 billion active parameters per token.
How does the recursive layer architecture work?
Inkling 2 Mini combines a Mixture-of-Experts approach with a design that runs specific layers repeatedly. The middle 12 of the decoder’s 24 physical layers are processed three times using the same weights. As a result, the model makes a total of 48 layer visits per token without increasing the number of physical layers.
For each token, the model selects eight of the 128 experts available per layer. Its attention mechanism combines sliding-window layers with full-attention layers; the sliding window is up to 1,024 tokens long. During repeated passes, messages from the outputs of previous passes are also taken into account. This architecture changes how the model performs computation, but the sources do not announce any integration specifically for architecture or 3D software.
Runtime support, hardware, and license
Hugging Face says runtime support is available for Transformers, llama.cpp, MLX, and SGLang. For llama.cpp, 4-bit IQ4_NL and 8-bit Q8_0 options are available. The SGLang installation instructions specify an SM103 prerequisite; the DSpark draft model can also be used for speculative decoding.
The Decoder reports that the model is available under the Apache 2.0 license and can be run locally on Macs with at least 32 GB of memory. The same source also lists NVIDIA DGX Spark as a deployment option. These are details reported for specific systems; memory requirements and performance on different hardware and software configurations need to be verified separately.
In the test comparison shared by Hugging Face, Inkling 2 Mini is shown leading in 32 of the 48 benchmarks compared with Gemma 4 31B, Muse Glimmer, Qwen3.8-27B, and Nemotron 3.5 Super VL. This reflects the evaluation results reported in the announcement; it does not guarantee performance on every workload.
A separate research version allows the number of loops to be changed. According to Hugging Face, reducing the number of passes to 1 or 2 may speed up inference, but could lead to some decline in output quality. The source also notes that this version has not been extensively evaluated.
What does this mean for architecture and visualization teams?
Image input support opens the door to experiments such as feeding architectural references or scene images to the model and asking questions about their content. The long context window also makes it possible to work with large text inputs. However, the announcements do not show a direct connection to BIM, modeling, or rendering applications, so these features should not be treated as ready-made design automation.
Offices and visualization teams should first check the license terms, compatibility with their existing hardware, and the memory requirements of their chosen deployment method. If using SGLang, they should take into account the SM103 prerequisite noted in the source. Before bringing the model into production workflows, teams should test visual interpretation and text-based tasks using examples from real projects.
Sources
1 sourceSource texts are not republished; short quotes are marked, everything else is our own summary and commentary.
Inkling 2 Mini’s ability to accept image input and its local deployment options could allow architecture offices and visualization teams to try auxiliary tasks such as reference analysis in their own environments. However, the sources do not show direct integration with design software, so the model’s role in the production pipeline should be determined through careful testing for now.
For teams in Turkey, it is important to evaluate the Apache 2.0 license, their own terms of use, and hardware compatibility together. Macs with at least 32 GB of memory and NVIDIA DGX Spark are the options mentioned in the source; performance on other systems should be measured separately. Any potential speed gains from reducing the number of loops should also be assessed alongside quality.
Frequently asked questions
Does Inkling 2 Mini support image inputs?
Yes. The model accepts text and image inputs and offers a 1 million-token context window.
Which systems can run Inkling 2 Mini locally?
The Decoder lists Macs with at least 32 GB of memory and NVIDIA DGX Spark. Hugging Face also reports support for Transformers, llama.cpp, MLX, and SGLang.
What is Inkling 2 Mini’s license?
The Decoder reports that the model is available under the Apache 2.0 license.



