Meta Releases Muse Glimmer, a Laptop-Ready AI Model, as Zuckerberg Pushes for Looser Open-Source Rules

Meta released Muse Glimmer on Monday, a 30 billion parameter open-weight AI model designed to run on a Mac or PC with a single consumer graphics card.

CEO Mark Zuckerberg also said Meta plans to open source the weights of its more powerful Muse Spark 1.2 model soon.

The releases came alongside a 14-page essay from Zuckerberg calling for looser U.S. policies on open source AI, as Meta works to prove its costly superintelligence division can compete with closed labs such as OpenAI and Anthropic. At the same time, Chinese rivals including Moonshot, Alibaba, and DeepSeek continue to lead open-weight development.

A Compressed Model Built to Feel Instant

Meta’s research blog detailed the engineering tradeoffs behind fitting a 30 billion parameter model onto ordinary hardware. 

It uses 4-bit quantization to shrink the language model, paired with a lightweight “drafter” based on DFlash, a speculative decoding technique that proposes token blocks for the main model to verify in parallel, helping it generate responses faster. 

Mashable noted a practical limitation for anyone hoping to run Glimmer on an older laptop: Meta says it needs around 24GB of VRAM for good performance, typically requiring a high-end graphics card or a more powerful Apple M series chip.  

That slightly weakens Meta’s democratization pitch. Glimmer is free to download, but running it locally at a usable speed still comes with a significant hardware cost, unlike Google’s Gemma 4 12B, which runs easily with 16GB of VRAM.

Zuckerberg’s Policy Pitch Targets Distillation Rules

Reuters reported that Zuckerberg used the release to argue U.S. labs face a structural disadvantage against Chinese open-weight developers because American companies face additional restrictions on training data. 

He specifically called for policy changes around distillation, an allegation Chinese AI labs continue to face from the U.S.

The argument is notably self-interested: Meta built Glimmer using the same technique, transferring reasoning capabilities from the larger Muse Spark 1.2 into a smaller, shippable model.

The timing also reflects growing competition from Moonshot’s Kimi K3, Alibaba’s Qwen3.8-Max, and DeepSeek’s V4-Flash, with Chinese open-weight models increasingly matching top American systems on performance.

The Real Divide Isn’t Open Versus Closed, It’s Owned Versus Rented

Meta’s open-source framing is a sharper distinction worth sitting with: the company isn’t actually open-sourcing its best model, it’s open-sourcing a smaller distilled copy while keeping Muse Spark itself closed and firmly under company control. 

That’s a meaningfully different promise from Zuckerberg’s vision of distributing superintelligence so people can “direct it” themselves: the model users can actually own and run locally is, by design, a step below whatever Meta considers genuinely frontier-capable.

With Meta’s stock still roughly 10% down for the year, Monday’s dual announcement, a free model and a policy manifesto, reads as much like an attempt to restore investor confidence in its AI trajectory as a genuine stance on AI access.

The real test will be whether Muse Spark 1.2’s weights, once they ship, offer the same capabilities as Meta’s cloud-hosted version or a deliberately hobbled one.

Source: Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device