Meta launches Muse Glimmer: What it means for the open-weight vs closed AI debate | Technology News

The debut of Muse Glimmer marks Meta’s partial return to the open-source approach to foundation model development since the high-profile exit of AI researcher Yann LeCun from the company last year. While Meta initially helped lead the open-source AI push with its Llama series of models released under a custom license, the tech giant pivoted towards developing closed, proprietary LLMs after falling behind OpenAI and Anthropic in the AI race.

It comes as Silicon Valley remains divided over a growing challenge from a wave of new Chinese open-weight AI models such as Moonshot AI’s Kimi K3 and Alibaba’s Qwen, which are not only closing the performance gap with US-developed frontier models but also reportedly undercutting them on cost.

In a post on X, CEO Mark Zuckerberg wrote, “Meta is a strong supporter of open source and I’m proud of these releases.”

“Foreign ⁠labs currently hold several advantages since American labs have to comply with many additional restrictions on training data,” Zuckerberg said in a longer blog post, referring to open-weight AI models. “US policy must reduce this additional ​friction if we want American open source models to lead over time,” he further said, ​adding that restricting ⁠access to foreign open-source models was not an effective solution.

Today we’re also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally. Soon we’ll also release the weights for Muse Spark 1.2, our latest foundation model. Meta is a strong supporter of open source and I’m proud of these releases. Congrats…

— Mark Zuckerberg (@finkd) August 10, 2026

In the coming months, Meta plans to release an open-weight version of Muse Spark 1.2 and a new cutting-edge AI model, internally called ‘Watermelon’, that is said to be more powerful than Muse Spark. However, it is unclear if the latter model would be open or closed.

Under the hood

In order to develop an SLM that overcomes the compute constraints of local hardware, Meta said that it set out to build a model with a compact architecture. It trained Muse Glimmer on the outputs of a much larger, ‘teacher’ model via distillation – a commonly used AI/ML technique that has grown controversial in recent times. The data used to train Muse Glimmer was a mix of synthetic and organic data spanning more than 100 languages.

During post-training, Meta said that it combined supervised fine-tuning with a mix of on-policy distillation and reinforcement learning across general, reasoning, coding, and agentic domains. The company further said that it utilised inference optimisation techniques such as quantisation in order to meet the latency requirements for running the model locally on-device.

In terms of safety, Meta said that Muse Glimmer was evaluated against standards laid out in its own safety framework for assessing open-weight models before their release.

‘Built to run on consumer hardware’

Meta said it applied two techniques to ensure that Muse Glimmer can run at practical speeds on consumer hardware without trade-offs such as compromising on the quality of model outputs.

While a 30-billion parameter model would typically require over 55GB of memory, Meta was able to lower the memory requirement to under 20GB by compressing the model weights to approximately 4-bit precision, ensuring minimal to no degradation on agentic tasks.

Muse Glimmer is made up of two key parts: the main model and a ‘drafter model’ that is used to process entire blocks of tokens at once. These tokens are then verified by the main model, enabling it to generate text significantly faster than standard token-by-token generation by other language models.

Local use cases, performance of Muse Glimmer

Muse Glimmer is a foundational SLM that comes with long-horizon execution, precise tool calling, multimodal understanding, long-context memory, and instruction following. Bringing these capabilities together, the model can be used to power and deploy autonomous AI agents that can manage schedules, draft messages, organise files, etc.

Based on its performance across a range of benchmarks, Meta claimed that Muse Glimmer’s strengths include end-to-end agentic task completion, reliable tool use, multi-step reasoning, multimodal input and reasoning, among others. It is further designed to work with existing developer tools, with support for llama.cpp, MLX, and ExecuTorch integrations available in the coming days.

Among models of its own class and size, Muse Glimmer beat Google’s 31-billion-parameter Gemma model and Alibaba’s 27-billion-parameter Qwen model across agentic, coding, multimodal, safety, and reasoning benchmarks.

To measure its speed on consumer hardware, Meta said that it ran the model on devices equipped with the Macbook M4 Max chip, M5 Max chip, and an Nvidia RTX-5090 GPU. The results showed that the model is fast enough for fluid conversation and real-time agent interaction without depending on cloud infrastructure and network access.