← All stories
● Covered by 2 sources · 2 reportsMedium impact2 positive

Inkling 2 Mini Model Released, Designed for Local AI Setups with Looped Transformer

🔄 Updated 20h ago
New to BrevFeed? We gather this story from every outlet covering it into one summary — ranked by real-world impact, not just the latest headline — so you never miss what matters. What is BrevFeed? →

Key points

  • Inkling 2 Mini is a multimodal Mixture-of-Experts model with 26.4B total parameters.
  • It uses a looped transformer architecture, repeating middle layers three times for increased reasoning.
  • The model supports a 1M-token context and accepts both image and text inputs.
  • Designed for local AI setups, it runs on consumer laptops and includes NVFP4 and MXFP8 expert weights.
  • Outperforms open-weight models under 100B parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%).

Overview of Inkling 2 Mini

Thinking Machines Lab has introduced Inkling 2 Mini, a new multimodal Mixture-of-Experts (MoE) model. This model has approximately 26.4 billion total parameters, with 3.4 billion active parameters per token. It is designed for local AI setups, making it suitable for deployment on consumer laptops.

Looped Transformer Architecture

A key feature of Inkling 2 Mini is its looped transformer architecture. This design reuses certain layers, specifically the middle layers, by running them three times for each token. This technique effectively increases the model's logical parameter count and allows for more extensive reasoning without explicitly generating a chain of thought.

The looping mechanism contributes to improved reasoning quality and efficiency, particularly for tasks requiring complex thought processes, such as agentic coding. This approach also helps maintain monitorability, allowing for oversight of the model's operations.

Performance and Capabilities

Inkling 2 Mini supports a 1-million-token context size and can process both image and text inputs. The model is shipped in a mixed-precision format, utilizing NVFP4 and MXFP8 expert weights. It also includes a DSpark drafter to enhance speed with minimal memory overhead.

On benchmarks, Inkling 2 Mini leads open-weight models with fewer than 100 billion total parameters. It achieved 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2. The model is noted to run faster than other models with comparable performance.

Deployment and Support

The model is provided with support for transformers, Llama App (llama.cpp), and MLX. Additionally, SGLang offers day-zero inference support for production deployments on datacenter hardware. This broad support aims to facilitate its integration into various environments.

✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →

The daily brief

One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.

One email a day. Unsubscribe in one click, any time.

Today's brief

Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.

~4 min · 3 stories · Nov 21

▶ Play today's brief Listen on Spotify

New every morning, and the back catalogue is archived by date.

How outlets covered it

Inkling 2 Mini, a new 26.4B parameter AI model, has been released with open weights, designed to run on consumer laptops. It uses a looped transformer architecture to improve reasoning quality and efficiency, outperforming other open-weight models under 100B parameters on benchmarks like DeepSWE v1.1 and Vals Index v2.2. The model's architecture allows for longer reasoning without explicit chain-of-thought, while maintaining monitorability.

Thinking Machines Lab has released Inkling 2 Mini, a new multimodal Mixture-of-Experts (MoE) model designed for local AI setups. It features a looped transformer architecture that reuses layers to effectively increase logical parameter count, supporting a 1M-token context and accepting image and text inputs.