Meta Reopens Its Open-Source AI Strategy with Muse, Glimmer, and Spark 1.2

Meta is back into open source!

On Monday, the company, with a market capitalization of approximately $1.5 trillion, released the underlying parameters of its newly open AI model, Muse Glimmer, allowing developers to download and modify them directly.

Meta also stated that it will release the model weights for its more powerful Muse Spark 1.2, the latest foundational model from Meta, over the coming weeks.

In response, Zuckerberg said that Meta has always been firmly committed to open source, and he is very proud of these releases.

Zuckerberg also published an article on Meta’s official website, proposing to provide billions of people with free and powerful AI, aiming to empower individuals and balance the power of large institutions.

Article URL: https://www.meta.com/thefutureisforeveryone/

In a passage clearly targeting companies like Google, Anthropic, and OpenAI, he wrote: Most other labs are focused on building AI for businesses, governments, or other institutions. If these labs ultimately become dominant, the balance of power will shift toward large institutions rather than individuals.

These commitments mean that Meta is returning to its previous open AI approach, which previously set it apart from competitors.

Earlier this year, Meta withheld the underlying weights of Muse Spark citing security concerns. As competition among leading AI companies intensifies, a growing central question has emerged: To what extent should the most powerful AI technologies be made open?

Nevertheless, Zuckerberg's renewed embrace of an open approach has been met with widespread praise in the comments section.

Yann LeCun rarely offered praise to his former employer.

Other netizens commented, "This is exactly what we needed! Zuckerberg has really made a comeback."

We love OpenAI (dog head)

Muse Glimmer is Meta's first open-weight model since Llama 4 and the first model released by Meta under the Apache 2.0 license.

Previous Llama series used a custom license with restrictions on certain commercial use cases. Apache 2.0 is much more permissive, allowing enterprises to deploy directly, continue fine-tuning, develop derivative models, and integrate them into their own products.

Muse Glimmer is distilled from Muse Spark, with a total of approximately 29.6 billion parameters, including an 1.8-billion-parameter visual encoder. The model supports text and image inputs, features a 128K context length, and is optimized for multi-step reasoning, tool invocation, and fault recovery, designed as an agent foundation for personal devices.

Traditional cloud-based agents require continuous uploading of files, messages, and work contexts; the longer the task, the more frequent the calls, and the higher the token costs. After deployment locally, data can remain on the device, making latency and inference costs easier to manage.

Hardware requirements still exist. The BF16 weights of Muse Glimmer are close to 60 GB, making it difficult for ordinary computers to handle directly. Meta’s 4-bit quantized version compresses the language model to under 20 GB, enabling it to run in environments with 24 GB or 32 GB of VRAM.

所谓“可运行于个人电脑”，更准确地说，是指配备高-end Mac or GPUs such as the RTX 5090.

In terms of parameter efficiency, Muse Glimmer has delivered an impressive performance.

Artificial Analysis scores 35 in intelligence, 21 points higher than Llama 4 Maverick. It is on par with the 36-point score of Kimi K2.5 is close to, but slightly below, Qwen3.6 27B and Ling 3.0 Flash, both of which scored 38.

In Meta’s published Agent evaluation, Muse Glimmer scored 75.5 on MCP Atlas, surpassing Qwen3.6 27B’s 62.5; it achieved 74.6 on DeepSearch QA, also slightly higher than the latter’s 71.1. In the tool-use benchmark 𝛕3-Banking, it scored 23.5, leading models in its class.

Independent evaluations also revealed shortcomings. Muse Glimmer scored 953 Elo on GDPval-AA v2, below the human baseline of 1000 and trailing Qwen3.6 27B’s score of 1141. On the AA-Omniscience evaluation, Muse Glimmer had a hallucination rate of 82%, compared to 49% for Qwen3.6 27B. On Terminal-Bench 2.1, Muse Glimmer scored 52%, again lower than Qwen3.6 27B’s 61%.

Thus, Muse Glimmer is well-suited for local tool invocation and process execution, but for tasks requiring high accuracy, it should still be combined with retrieval and human review.

In Zuckerberg’s view, the core of the AI competition boils down to two questions: Who will achieve superintelligence, and what will people do with it?

He aims to give "personal superintelligence" to billions of users, enabling AI to assist in daily matters such as health, career, finance, interests, and relationships.

This position ties open models to individual autonomy and directly challenges the proprietary approaches of companies like OpenAI and Anthropic.

Zuckerberg opposes using security risks to justify centralization of power.

He believes that entrusting the most powerful AI to a small number of institutions inherently creates new power risks. Open models allow more developers to review them, making vulnerabilities easier to detect and fix, and enabling individuals to customize AI according to their needs.

He also defended model distillation.

OpenAI and Anthropic have recently accused Chinese companies of using outputs from American proprietary models to train their own models. Zuckerberg emphasized that people should uphold the principle of "learning from observable information." He also opposes restricting overseas open models, advocating that American open models should compete to become the best in the world.

However, Meta did not avoid all security issues.

Zuckerberg proposed that Meta’s independent directors be responsible for approving the security standards required for model openness; the company could also provide intermediate training checkpoints to governments to enable earlier detection of risks.

Meta's return to open weights at this time is also driven by practical pressures.

China's open models are rapidly closing the gap with America's proprietary flagship models. DeepSeek . Kimi Continue improving Qwen's performance and expanding its reach through low pricing, customization, and on-premises deployment. If U.S. model companies continue to tighten weight restrictions, the global developer ecosystem may accelerate its migration to other providers.

Meta needs to win back these developers. Muse Glimmer lowers the barrier to local deployment, while Muse Spark 1.2 raises the capability ceiling—both models support personal devices, enterprise private deployment, and cloud services.

Openness also creates demand for Meta’s computing investments. This year, Meta plans to invest up to $145 billion in AI infrastructure while preparing its cloud computing business. While model weights can be free, inference, hosting, and development tools can still generate revenue. Openness expands the ecosystem, enabling cloud services to drive commercialization.

The $1 billion data center community fund announced the same day also supports this strategy. Meta needs to build more data centers while addressing local community concerns about electricity, water resources, and land use. Technology openness, policy advocacy, and infrastructure expansion are thus interconnected.

Muse Glimmer is not yet sufficient to reshape the model landscape. Its significance lies in Meta rejoining the open-weights camp and delivering a usable product designed for local agents.

If Meta opens Muse Spark 1.2 as planned, competition between open-source and closed-source models will intensify further.

Next, the industry must answer three questions: Which capabilities can be opened up, who bears the cost, and who sets the rules.

Reference link:

https://x.com/ArtificialAnlys/status/2086916150278111551

https://www.ft.com/content/4e3957f8-ea7c-4c46-a3de-cdce8e526878?syn-25a6b1a6=1

https://www.bloomberg.com/news/articles/2026-08-10/meta-releases-muse-glimmer-ai-model-people-can-run-on-their-laptop?srnd=phx-technology

This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), authored by someone interested in AI.