Skip to content
important New models

Thinking Machines Lab releases Inkling 2 Mini model with looped transformer architecture

clearly official source

Thinking Machines Lab has released the multimodal Inkling 2 Mini model (Mixture-of-Experts, 26B/3B parameters, 1M token context), intended for local deployment. According to the company, it is comparable to the Qwen 3.8 27B model.

Thinking Machines Lab has released the Inkling 2 Mini model on the Hugging Face platform — a multimodal (text and images) Mixture-of-Experts model with a "looped transformer" architecture. The model has 24 physical decoder layers, but the middle 12 layers are repeated three times in a loop, so it effectively corresponds to 48 logical layers. The total number of parameters is approximately 26 billion, with approximately 3 billion active per token. The router in each layer selects 8 experts out of 128 available. The model supports a context of up to 1 million tokens and is delivered in a mixed-precision format with NVFP4 and MXFP8 expert weights; for further speedup it also includes a so-called DSpark drafter.

According to the company, between individual loop passes the model passes information using "cross-pass messages" and controlled gating, which decides how much of the previous pass should be retained and how much should be replaced by new input. Alongside the main Inkling 2 Mini model, the company also released a research variant, TML-d24-varloop-research, which allows changing the number of loop passes (1 to 3) — fewer passes mean faster inference, but according to the company potentially lower output quality; this variant has been tested only to a limited extent.

According to the company, Inkling 2 Mini is the best in its class in the Artificial Analysis Intelligence Index v5, and compared to the Gemma 4 31B, Muse Glimmer, Qwen3.8-27B, and Nemotron 3.5 Super VL models, it leads in 32 of 48 tracked benchmarks, including coding, agentic search, and visual tasks. According to the company, the model leads in the FrontierCode v2.0 benchmark, ties with the Qwen3.8-27B model in Terminal-Bench 4.0, and lags behind it in AutomationBench-AA v1.0.6. Support is provided in the transformers, llama.cpp, and MLX libraries for local deployment, and in SGLang for production deployment on data center hardware.

What changed

Why it matters

The model is aimed at local deployment (smaller size, support in llama.cpp and MLX), making it relevant for developers and companies who want to run AI models without relying on cloud APIs. The combination of a looped architecture and Mixture-of-Experts promises higher effective capacity without a proportional increase in the number of physical layers and memory requirements, which could reduce hardware requirements compared to models with comparable performance. The comparative figures and qualitative assessments come from the model's publisher; the source does not mention independent verification.

Release card

Inkling 2 Mini

Thinking Machines Lab

open weights
Specifications
approximately 26 billion parameters in total (Mixture-of-Experts), approximately 3 billion active per token
Context
up to 1 million tokens
Inputs
text and images as input, text as output
Availability
The model is available for download on the Hugging Face platform (thinkingmachines/Inkling-2-Mini) with support in the transformers, llama.cpp, and MLX libraries for local deployment, and in SGLang for production deployment on data center hardware.
Documented measurements
According to the sources, it is suitable for
  • local deployment thanks to its compact size and context of up to 1 million tokens
  • coding tasks, where the model leads in the FrontierCode v2.0 benchmark
  • working with both text and images thanks to multimodal inputs
Documented limits
  • in the AutomationBench-AA v1.0.6 benchmark it lags behind the Qwen3.8-27B model
  • the research variant with a variable number of loop passes has been tested only to a limited extent

The company compares the model with the Gemma 4 31B, Muse Glimmer, Qwen3.8-27B, and Nemotron 3.5 Super VL models and, based on its own informal tests, considers it at least comparable to the Qwen 3.8 27B model.

The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →

Two audiences, two different impacts

What this means

01

For individuals

Developers and technically oriented users can try a new open-weight model for local use with support in transformers, llama.cpp, and MLX, including multimodal input and long context.

What to do Try the Inkling 2 Mini model locally via transformers, llama.cpp, or MLX for a specific task and compare it with your current solution.
More practical updates →
02

For a business

Companies considering local deployment of AI models (without sending data to the cloud) gain another open-weight alternative with lower active-parameter requirements; according to the manufacturer, it has competitive performance compared to models of a similar class, but independent verification is still lacking.

Development
What to decide Independently verify the performance and license of the Inkling 2 Mini model before any deployment in internal tools.
More business impacts →
Inkling 2 Mini local AI looped transformer mixture-of-experts open model Thinking Machines Lab

Check the original

Event sources

clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.

1
Hugging Face Blog primary source · first detected Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup