🤖 AI Summary
Today, the release of Inkling 2 Mini marks a significant advancement in AI model design, offering a powerful looped transformer model that can run efficiently on standard laptops. This innovative architecture processes each token through its middle layers three times, enhancing reasoning capabilities and yielding higher accuracy with fewer tokens—particularly beneficial for complex tasks like agentic coding. Inkling 2 Mini outperforms every comparative model under 100 billion parameters in key benchmarks such as DeepSWE v1.1 and Vals Index v2.2, while maintaining a lightweight footprint of around 20GB.
The model employs a Mixture-of-Experts approach with 26.4 billion total parameters, where only 3.4 billion are activated per token through looping. This functionality not only fosters longer and clearer reasoning chains without losing legibility but also improves the model's monitorability. The introduction of a variable-loop variant, designed for research comparisons, shows that more passes enhance performance on serial reasoning tasks without negatively impacting the model’s integrity in terms of agentic sabotage. With open weights available on Tinker, Inkling 2 Mini is positioned as a cutting-edge tool for researchers and developers, promising advancements in AI-driven reasoning and application in various practical coding scenarios.
Loading comments...
login to comment
loading comments...
no comments yet