Meta is back to open source again!
On Monday, the company, with a market value of about $1.5 trillion, released the underlying parameters of its new open AI model Muse Glimmer, which developers can directly download and modify.
Meta also said that in the coming weeks, it will also open the model weights of its more powerful Muse Spark 1.2, Meta's latest foundation model.
In response, Zuckerberg said: Meta has always firmly supported open source, and he is very proud of these releases.
Zuckerberg also published an article on Meta's official website, proposing to provide free and powerful AI to billions of people, hoping to empower individuals and counterbalance the power of large institutions.
Article URL: https://www.meta.com/thefutureisforeveryone/
In a statement clearly aimed at companies like Google, Anthropic, and OpenAI, he wrote: Most other labs are focused on building AI for enterprises, governments, or other institutions. If these labs end up taking the lead, the balance of power will tilt more toward large institutions rather than individuals.
These commitments mean that Meta is returning to the open AI path it previously used to differentiate itself from competitors.
Earlier this year, Meta did not open the underlying weights of Muse Spark, citing safety concerns. As competition among leading AI companies intensifies, an increasingly central question has become: to what extent should the most powerful AI technology be open?
But regardless, Zuckerberg's re-embrace of the open path this time has been met with almost unanimous praise in the comment section.
Yann LeCun rarely gave his former employer a round of applause.
Some netizens also said, "This is so on point! Zuckerberg has really made a comeback."
"We love open AI (dog head)"
Meta Replays the "Open Source Card"
Muse Glimmer is Meta's first open-weight model since Llama 4, and also Meta's first model released under the Apache 2.0 license.
Previous Llama series used a custom license with restrictions on certain commercial scenarios. 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 parameter count of about 29.6 billion, including a vision encoder with about 1.8 billion parameters. The model supports text and image inputs, has a 128K context, and focuses on enhancing multi-step reasoning, tool calling, and fault recovery, positioned as the Agent foundation for personal devices.
Traditional cloud-based Agents require continuous uploading of files, messages, and work context. The longer the task, the more calls, and the higher the token bill. After deploying locally, data can stay on the device, and latency and inference costs are easier to control.
Hardware barriers still exist. Muse Glimmer's BF16 weights are close to 60GB, making it difficult for ordinary computers to handle directly. Meta's 4-bit quantized version compresses the language model to under 20GB, allowing it to run in environments with 24GB or 32GB of VRAM.
The so-called "runs on personal computers" more accurately means high-end Macs or PCs with graphics cards like the RTX 5090 can run it.
In terms of parameter efficiency, Muse Glimmer delivers a decent performance.
Artificial Analysis gives it an intelligence index of 35, 21 points higher than Llama 4 Maverick. It is close to
Kimi
K2.5's 36, and slightly lower than Qwen3.6 27B and Ling 3.0 Flash, both of which score 38.
In Meta's announced Agent evaluations, Muse Glimmer scores 75.5 on MCP Atlas, higher than Qwen3.6 27B's 62.5; DeepSearch QA is 74.6, also slightly higher than the latter's 71.1. In the tool-use benchmark 𝛕3-Banking, it scores 23.5, leading models in its class.
Independent evaluations also expose shortcomings. Muse Glimmer scores 953 Elo on GDPval-AA v2, below the human baseline of 1000, and also behind Qwen3.6 27B's 1141. Its hallucination rate on the AA-Omniscience evaluation is 82%, while Qwen3.6 27B is 49%. On Terminal-Bench 2.1, Muse Glimmer scores 52%, also lower than Qwen3.6 27B's 61%.
Thus, Muse Glimmer is suitable for local tool invocation and workflow execution, but for high-accuracy knowledge tasks, it still needs to be combined with retrieval and human review.
Zuckerberg's Battle for AI Power Narrative
In Zuckerberg's view, the core of AI competition can be reduced to two questions: who gets superintelligence, and what people will do with it.
He hopes to put "personal superintelligence" into the hands of billions of users, letting AI participate in daily matters such as health, career, finance, interests, and relationships.
This proposition ties open models to personal autonomy and directly challenges the closed-source approaches of companies like OpenAI and Anthropic.
Zuckerberg opposes justifying the concentration of capabilities with safety risks.
He believes that placing the most powerful AI under the control of a very small number of institutions itself creates new power risks. Open models can be inspected by more developers, vulnerabilities are more easily exposed and fixed, and individuals can customize AI as needed.
He also defends model distillation.
OpenAI and Anthropic have recently repeatedly accused Chinese companies of using outputs from US closed-source models to train their own models. Zuckerberg emphasizes that people should retain the principle of "learning from observable information." He also opposes restricting overseas open models, advocating that US open models should rely on competition to become the best globally.
However, Meta has not avoided all safety issues.
Zuckerberg proposes that Meta's independent directors be responsible for approving the safety standards required for model release; the company can also provide intermediate training checkpoints to the government to identify risks earlier.
Behind Open Source, There's Also the Ecosystem and Computing Power Business
Meta's return to open weights at this time also faces practical pressures.
Chinese open models are rapidly narrowing the gap with US closed-source flagships.
DeepSeek
,
Kimi
, and Qwen continue to improve performance and expand influence through low prices, customizability, and local deployment. If US model companies continue to tighten weights, the global developer ecosystem may accelerate migration to other providers.
Meta needs to win back these developers. Muse Glimmer lowers the barrier for local deployment, while Muse Spark 1.2 raises the capability ceiling; the two models can cover personal devices, enterprise private deployment, and cloud services.
Openness can also create demand for Meta's computing power investments. Meta plans to invest up to $145 billion this year in AI infrastructure, while also preparing its cloud computing business. Model weights can be free, but inference, hosting, and development tools can still generate revenue. Openness expands the ecosystem, and cloud services commercialize it.
The $1 billion data center community fund announced the same day also serves this layout. Meta needs to build more data centers and also alleviate local community concerns about electricity, water resources, and land use. Technology openness, policy advocacy, and infrastructure expansion are thus connected in a line.
Muse Glimmer is not yet enough to rewrite the model landscape. Its importance lies in Meta re-joining the open-weight camp and offering a usable product for local agents.
If Meta releases Muse Spark 1.2 as planned, competition between open and closed source will further intensify.
Next, the industry must answer three questions: which capabilities can be open, who bears the cost, and who sets the rules.
Reference links:
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 Intelligence" (ID: almosthuman2014), authored by those who follow AI.
