Signal

Artificial Analysis says Inkling 2 Mini, with 5.4B active parameters, matches Qwen3.8 27B (xhigh) on its Intelligence Index with ~5x fewer active parameters

First reported by Artificialanalysis ·

The signal ●●○○ Compiled by AI from Artificialanalysis and Techmeme
Why you might care

Running capable AI models on personal hardware becomes feasible and significantly more affordable.

What happened

Thinking Machines has released Inkling 2 Mini, an open-weights looped transformer model that performs comparably to larger models like Qwen3.8 27B on the Artificial Analysis Intelligence Index. The model boasts 26 billion total parameters, with 5.4 billion active parameters per token, allowing it to run on consumer hardware like a MacBook. Despite using significantly fewer active parameters, Inkling 2 Mini achieves intelligence scores similar to Qwen3.8 27B and Nanbeige4.3-33B, while requiring substantially less compute and costing considerably less per task. It processes approximately 152K output tokens per Intelligence Index task on average, uses about 2.8 times less compute than Qwen3.8 27B, and is approximately 6 times cheaper and 4 times faster. The model supports text and image inputs with a 1 million token context window and is licensed under Apache 2.0.

What it means

The development of Inkling 2 Mini, particularly its efficient use of active parameters through looped layers, signals a growing trend towards making high-performance LLMs accessible on edge devices and personal computers. This approach contrasts with the massive parameter counts of many state-of-the-art models, suggesting a potential shift in architectural design focused on computational efficiency rather than sheer scale. The Pareto frontier placement indicates that Thinking Machines is successfully balancing intelligence with resource requirements, a key challenge in democratizing AI capabilities.

This advancement directly impacts developers and researchers who can now experiment with and deploy sophisticated AI models without requiring extensive cloud infrastructure or specialized hardware. The reduction in cost and increase in speed per task for comparable intelligence scores could accelerate innovation cycles and open new use cases for AI applications. Competitors will likely face pressure to optimize their own models for similar efficiency gains, potentially leading to a broader market shift towards more accessible and cost-effective AI solutions.

AI-written summary. May contain errors.

Chips