TML releases Inkling 2 Mini, an open-weights 26.4B-parameter transformer that loops middle layers three times per token for better reasoning with fewer tokens
First reported by Thinkingmachines ·
You can now run advanced reasoning models locally on your laptop that previously required significant cloud resources.
Thinking Machines Lab (TML) has released Inkling 2 Mini, an open-weights transformer model with 26.4 billion parameters. This model is designed with a novel looped transformer architecture, where middle layers are processed three times per token. This design enables the model to perform enhanced reasoning with fewer tokens, demonstrating superior performance on agentic coding tasks. On benchmarks such as DeepSWE v1.1 and Vals Index v2.2, Inkling 2 Mini has reportedly surpassed all other open-weights models with fewer than 100 billion parameters. The model's architecture, including its NVFP4 and low-bit quantization, allows it to have a memory footprint of approximately 20GB, making it capable of running on consumer laptops. TML has made Inkling 2 Mini available on Tinker and Hugging Face.
The key innovation in Inkling 2 Mini is its looped transformer architecture, which reruns middle layers multiple times per token. This approach enhances the model's reasoning capabilities without a proportional increase in computational cost or parameter count. The improved reasoning is particularly beneficial for tasks requiring multi-step thinking, such as agentic coding, allowing these complex operations to be performed more efficiently and with greater accuracy on less powerful hardware.
This release signals a potential shift towards more capable, yet accessible, on-device AI. By optimizing for local execution and reasoning depth, Inkling 2 Mini challenges the trend of ever-larger, cloud-bound models. Its performance on coding and agentic tasks suggests a future where sophisticated AI assistance can be integrated into personal devices for offline use, impacting how developers and users interact with AI tools.
AI-written summary. May contain errors.