Small Language Models

Supersonic Labs has released Julia 1, an open decision model with 144.3 million parameters that is small enough to run on a standard CPU. Unlike large language models, decision models are built for making choices and taking actions rather than generating text. The low hardware requirement makes the model accessible to developers without expensive GPU infrastructure, which is drawing attention in the AI community.

AWS Strands Labs has released Strands Decider 2B, a new open source model designed to make quick decisions between options. According to the announcement, the model picks among choices in roughly 115 milliseconds, making it suited to latency-sensitive applications where larger language models would be too slow. The release adds to the growing set of small, specialized open models aimed at specific tasks rather than general-purpose reasoning.

A quantized 27-billion-parameter language model is reported to match frontier AI models on a single task from a coding benchmark. The narrow, specific nature of the claim makes it more believable than sweeping benchmark-superiority claims, but it also means the result says little about overall performance. Readers are debating how much weight such partial benchmark results deserve in judging open and smaller models.

AI startup PrismML says it has adapted its small language models to run on smart glasses powered by Qualcomm chips. The company's broader aim is open-weight AI that runs entirely on devices, making better use of the computing power hardware already has rather than relying on the cloud. The move highlights growing interest in compact, on-device AI models for wearables.

Developers are weighing whether to switch their apps and services from pay-per-use AI APIs to flat subscription models as API costs for large language models keep climbing. Supporters of subscriptions say predictable pricing protects margins and simplifies billing for users, while critics argue usage-based pricing is fairer and subscriptions can lead to losses when heavy users consume more AI compute than they pay for. The debate has split developer communities, with many sharing cost breakdowns and real-world examples of both approaches.