Wink Pings

26B-Parameter Model Runs on 32GB Mac With 79.2% SWE-bench Score: Thinking Machines Unveils Inkling 2 Mini

Thinking Machines has launched Inkling 2 Mini, a 26.4B-parameter looped MoE transformer with only 3.4B activated parameters. It can run on a 32GB Mac or a single DGX Spark, and scores 79.2% on SWE-bench Verified. This high-performance small model makes local AI execution a practical possibility.

Thinking Machines has released Inkling 2 Mini, a 26.4B-parameter looped MoE transformer with merely 3.4B activated parameters. It can run on a 32GB RAM Mac or a single DGX Spark, and achieves a 79.2% score on SWE-bench Verified.

Let's break down a few key terms first. A looped transformer reuses the same set of layers cyclically, which reduces total parameter count while increasing computational depth. MoE, short for Mixture of Experts, only activates a portion of experts during each inference run. That explains why even with 26.4B total parameters, its actual computation load is equivalent to a 3.4B-parameter model. This design drastically cuts inference costs while fully preserving the model's core capabilities.

Running smoothly on a 32GB Mac means the barrier for local deployment is extremely low. You do not need an A100 cluster — a regular MacBook Pro is enough to run an AI with coding performance close to that of top-tier large models. SWE-bench Verified is a benchmark that evaluates AI's ability to solve real-world GitHub issues, consisting of 500 problems that require the model to generate valid code patches that pass all associated tests. The 79.2% score is very impressive, nearly matching the performance of some leading closed-source large models.

Thinking Machines was founded by former OpenAI CTO Mira Murati, and Inkling 2 Mini is the company's second public model release. The signal sent by this launch is unambiguous: the future direction points to small models, high efficiency, and local execution. While the rest of the industry is still competing to build trillion-parameter giant models, this development path is well worth close attention.

Source link: https://www.marktechpost.com/2026/11/21/thinking-machines-inkling-2-mini-looped-transformer/

发布时间: 2026-11-21 20:13