Inkling 2 Mini
The Inkling 2 Mini model has been released, a compact and powerful machine learning model capable of running on a standard laptop. This looped transformer, which repeats the middle layers of the model three times for every token, enables the model to perform more reasoning and reach the correct answer with fewer tokens, particularly in tasks requiring strong reasoning like agentic coding.
Inkling 2 Mini outperforms other open-weights models in benchmarks such as DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%) despite having fewer than 100 billion parameters. It runs faster than similarly performing models as well.
As a Mixture-of-Experts model, Inkling 2 Mini has 26.4 billion total parameters, with 3.4 billion active and 5.4 billion weight activations per token, accounting for looping. Looping allows the model to reason more thoroughly without writing down its reasoning, which is crucial for monitorability, or the ability to oversee and understand what the model is doing through its chain of thought.
The researchers measured the impact of looping on both performance and monitorability, finding that the no-CoT, or chain of thought, time horizon improved significantly with three loops, and performance on a benchmark of serial reasoning continued to improve with up to ten loops.
The model's chain of thought remained legible throughout despite more loops, and it did not perform better on tests of agentic sabotage as loops were added. Inkling 2 Mini's weights were trained and deployed natively in NVFP4, with the last four layers in MXFP8, making it memory-efficient at around 20GB. This allows the model to run on local devices such as consumer laptops and desktop AI computers.
The model's performance mirrors that of Inkling 2 Small, excelling on long-horizon agentic work and being the top open-weights model of similar size on most tasks.
The release of Inkling 2 Mini marks the first available models with looped architecture to researchers. Early testing by researchers at Stanford's Hazy Research Lab and Scaling Intelligence Lab showed Inkling 2 Mini's performance to be significantly better than other models, with higher intelligence per joule and intelligence per watt.
While Inkling 2 Mini may not be the best small model for multi-turn chat or instructions with many constraints, it shows strong capabilities in agentic tasks, search and tool use, and knowledge work benchmarks that require efficient use of search and tools.
Written by urgent.news from Hacker News's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.