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 (Thinking Machines Lab)
Today we are releasing Inkling 2 Mini, a fast, powerful model that's small enough to run on your laptop.
Today, TML released Inkling 2 Mini, a compact transformer model that can run on a laptop. This model features a unique looping architecture where the middle layers of the transformer are run three times for each input token. This looping enables the model to apply more reasoning and achieve better results with fewer tokens, particularly on tasks that require high-quality reasoning, such as agentic coding.
Inkling 2 Mini boasts fewer than 100 billion total parameters and outperforms other open-weights models with fewer parameters on benchmarks like DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%). The model is also faster than similarly performing models.
Inkling 2 Mini operates as a Mixture-of-Experts model, with 26.4 billion total parameters, 3.4 billion active, and 5.4 billion weight activations per token. Looping allows the model to reason longer without explicitly writing its reasoning, which is beneficial for monitoring purposes. The variable-loop variant of Inkling 2 Mini, released alongside the main model, demonstrates that performance improves with more loops up to a certain point, after which knowledge-intensive benchmarks do not see significant improvements.
Researchers at Stanford measured Inkling 2 Mini's intelligence per joule, finding that it scores 0.703 on NVIDIA DGX Spark, significantly higher than other models like Qwen3.6-35B-A3B (0.430), Gemma-4-26B-A4B (0.159), and Qwen3.5-27B (0.036). The model also runs efficiently on local devices, such as a MacBook Pro, allowing for the creation of applications like a playable flight simulator.
Despite its strengths in long-horizon agentic work and efficiency, Inkling 2 Mini does not excel in multi-turn chat or instructions with many constraints, trailing behind larger models in those areas.
Written by urgent.news from Techmeme's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.