Meta is reshuffling how it builds artificial intelligence once more, and this round the pitch is openness. The company says its large language models will now center on open weights, and it paired that message with something concrete: a freely available model named Muse Glimmer, plus a promise to publish the weights for its stronger system, Muse Spark 1.2, over the coming weeks.

Chief executive Mark Zuckerberg wrapped the shift in an essay that runs past 6,000 words, setting out how he believes AI systems should be built and governed from here. Part of the piece reads as a jab at rivals. OpenAI and Anthropic keep their models proprietary, and both have lobbied Washington for cover against two things they treat as threats: large-scale distillation, in which one model trains another, and the open-weight systems pouring out of Chinese labs. Meta is choosing the other side of that line.

Muse Glimmer is the lighter of the two systems. It holds 30 billion parameters and, by default, a context window that stretches to 128,000 tokens. Meta produced it by distilling Muse Spark, the bigger, more capable model it shipped earlier in the year, and built it to run locally, on a user's own hardware, instead of being served from the cloud or reached through an API. Its weights ship open under an Apache 2.0 license.

To grasp why this reads as a reboot, follow where Muse Spark began. Meta first showed it in April: a frontier-class model, closed and fully proprietary. That was the company's first major release after a hard reshuffle of its AI teams a year ago, and it marked a sharp turn from the open-leaning reputation it once carried. The zigzagging kept on. July brought Muse Spark 1.1 with the company's first paid tier, another break from old habits. Version 1.2 followed on August 5, bundled with Muse Code, a coding agent that works from the terminal.

Reactions to Muse Code have been mixed. Developers mostly find it trailing the best from Anthropic and OpenAI on raw capability, while holding its own on price, occupying the very niche that plenty of Chinese open-weight models already fill.

Muse Glimmer does not reach for that tier at all. Built for consumer GPUs, it sits below the frontier on purpose. Its real significance lies elsewhere: it feeds a rising push to move some of the computing onto local devices, so people and companies depend less on the big labs and spend less there too. For a company whose recent past is a run of reversals, leaning into that trend is one more correction of course. Whether this one holds is the question left open.

This article was rewritten from reporting by Ars Technica. Source: Ars Technica