Meta Releases Muse Glimmer, a 30-Billion-Parameter Open-Weight AI Model That Runs on a Single Consumer GPU

Дата публикации: 11-08-2026 07:11:27

Meta has released Muse Glimmer, a 30-billion-parameter open-weight AI model that runs on a Mac or PC with a single consumer graphics card.
Thank you for being a Ghacks reader. The post Meta Releases Muse Glimmer, a 30-Billion-Parameter Open-Weight AI Model That Runs on a Single Consumer GPU appeared first on gHacks.

Основное содержимое страницы с новостью.

Meta has released Muse Glimmer, a 30-billion-parameter open-weight AI model that runs on a Mac or PC with a single consumer graphics card. The model is available as a free download from Hugging Face under the Apache 2.0 license.

This allows commercial use, modification, and redistribution, without the restrictions found in some other open models. Muse Glimmer is distilled from Muse Spark 1.2, Meta's closed flagship model launched on August 5. It is aimed at agentic tasks like schedule management, file organization, local coding, and function calling.

CEO Mark Zuckerberg published a 14-page essay alongside the release arguing for distributed, open AI development over centralized systems.

How Meta Fit Muse Glimmer on Consumer Hardware

To fit a model this large onto consumer hardware, Meta used aggressive compression. At full floating-point precision, 30 billion parameters would need over 55 GB of memory, which is more than any consumer graphics card currently offers.

Meta reduced the weights to about 4-bit precision, bringing the model size down to under 20 GB. This leaves enough space for the KV cache, an image-processing encoder, and a speculative decoding drafter, all within a 24 GB or 32 GB memory limit.

The drafter uses DFlash, a small companion network that suggests blocks of tokens for the main model to check in parallel. Meta reports that decoding is 3.1 times faster on an NVIDIA RTX 5090, 1.8 times faster on an Apple M5 Max, and 1.5 times faster on an M4 Max, with output quality matching standard token-by-token generation.

Training took place in three phases. First, Meta used logit distillation from the larger Spark model. The second phase focused on longer-context and agent-heavy data. The final stage combined supervised fine-tuning, on-policy distillation, and reinforcement learning.

Meta says Muse Glimmer performs well for its size compared to Gemma4-31B and Qwen3.6-27B. Benchmarks include DeepSearch QA, MCP-Atlas, ?-Bench, and SWE-Bench, which test retrieval, multi-step agentic planning, and code writing and debugging.

Capabilities, Tooling, and Hardware Support

Muse Glimmer includes features for autonomous work beyond what the benchmarks measure. It can diagnose failed tool calls and retry them instead of stopping.

The model accepts mixed text and images, including screenshots and documents. It works with agent orchestrators like OpenClaw, supports adjustable reasoning effort, and covers over 100 languages.

Meta said local operation lets people use AI "anywhere, anytime, with or without an internet connection," and that queries processed on-device never leave the machine.

Optimized integrations for llama.cpp, MLX, and ExecuTorch are due within days, while Ollama, LM Studio, Unsloth, Together AI, Fireworks AI, and OpenRouter supported the model at launch.

Meta pointed larger-scale inference workloads to vLLM and SGLang, and hardware optimization work is underway with AMD, Arm, Dell, Intel, and NVIDIA.

Zuckerberg’s Case for OpenAI and the Competitive Backdrop

Zuckerberg framed the release in explicitly political terms. He wrote that instead of centralizing superintelligence, the industry should "distribute it widely and give every person the ability to direct it," and argued that US developers face regulatory disadvantages relative to Chinese competitors on training data usage and model distillation.

The competitive landscape adds weight to Zuckerberg's argument. Chinese companies like Moonshot, Alibaba, and DeepSeek have become leaders in open-weight AI, with models that rival top US systems. Meanwhile, the main models from OpenAI, Anthropic, and Google are still closed.

Security concerns have been part of the open-versus-closed debate this summer. In late July, NVIDIA launched the Open Secure AI Alliance to develop open security tools, after OpenAI models under evaluation breached Hugging Face's infrastructure.

Meta had a similar issue when Muse Spark 1.1 accessed a real company during testing because of a misconfigured evaluation environment.

Muse Glimmer will not be the only open model in Meta's lineup for long. Zuckerberg confirmed the company will also release the weights for Muse Spark 1.2 itself, which would put a frontier-class Meta model into open-weight territory, writing that "we will resume releasing some open source models soon." Meta's stock rose nearly 3% in premarket trading on the announcement.

This release marks a change in direction for Meta Superintelligence Labs, led by Alexandr Wang. The lab launched Muse Spark 1.1 on July 9 as the first Meta model with a price tag through the Meta Model API.

It then released the Muse Code terminal coding agent, another paid product, alongside the Meta One subscription for consumers. Muse Glimmer takes the opposite approach, being free and fully open.

Meta's benchmark results are self-reported, so real-world performance for Muse Glimmer will depend on what developers find as they use it. Meta has not given a firm date for releasing the Muse Spark 1.2 weights, only saying it will happen soon.

The hardware needed to run Muse Glimmer smoothly will also depend on how well the optimized integrations work once they are available.

Схожие новости

#	Наименование новости	Тональность	Информативность	Дата публикации
1	Запустится на ПК: Meta выкатила нейросеть Muse Glimmer и $1 млрд на дата-центры	0	10.37	10-08-2026
2	Meta Confirms One of Its AI Models Breached a Company During a Misconfigured Cyber Test	0	3.48	10-08-2026
3	Meta* выпустила офлайн-ИИ Muse Glimmer для Mac и ПК	0	14.54	11-08-2026
4	Meta Announces First AI Data Center, Prometheus, Coming Online in 2026 With More Superclusters Planned	0	6.28	16-07-2026
5	Telegram: 10 августа 2026	0	16.95	10-08-2026
6	Google Launches ProducerAI for DeepMind-Powered Music Creation	0	8.54	25-02-2026
7	Apple Chooses Google Gemini Over OpenAI for Future Siri and AI Features	0	5.87	13-01-2026
8	Apple Sues OpenAI Alleging Former Employees Stole Trade Secrets for AI Hardware Development	0	4.44	11-07-2026
9	Meta says its AI went rogue	0	10.99	06-08-2026
10	Meta’s new coding agent is cheap (but it’ll cost you your data).	0	16	07-08-2026

Классификация: Пресс-релизы. Схожих патентов: 0. Схожих новостей: 10. Тональность: 0. Информативность: 5.06. Источник: www.ghacks.net.