Zuckerberg defends the open strategy of Meta and distillation in an essay; the company also plans to open the weights of Muse Spark 1.2
To accompany the release of the open model Muse Glimmer (30B, Apache 2.0), Meta published an extensive essay by Mark Zuckerberg defending distillation and its open strategy against OpenAI and Anthropic; it also promised to open the weights of Muse Spark 1.2.
Meta accompanied the release of Muse Glimmer with an essay by CEO Mark Zuckerberg titled “The Future is for Everyone", over 6000 words long, published on 11 August 2026. The essay defends the open AI strategy of Meta and contrasts it with the approach of OpenAI and Anthropic, which develop closed models. In the text, Zuckerberg also openly defends distilling models from other labs — training a smaller model on the outputs of a model from another company — on the grounds that “it is possible to learn from everything that can be observed". According to the sources, this puts him at odds with OpenAI and Anthropic, which repeatedly describe distillation by Chinese labs as unauthorized use of their models; Dario Amodei, CEO of Anthropic, has long warned against the proliferation of open models at the level of the most powerful systems and advocates stricter export controls on China.
Muse Glimmer itself, which accompanies the essay, has 30 billion parameters, a context window of 128 thousand or more tokens, and is available under the Apache 2.0 license on Hugging Face and through Ollama. It is intended for agentic tasks running locally on a Mac or PC with a single consumer graphics card, including work with images thanks to its own 1.8-billion-parameter perception encoder. The model was distilled from the larger closed model Muse Spark and, according to Meta, outperforms Gemma4-31B from Google and Qwen3.6-27B from Alibaba on agentic tasks, while Qwen is better at controlling a computer desktop and terminal. It is the first open model from Meta Superintelligence Labs since Llama 4 in spring 2025.
According to the Wall Street Journal, cited by the source The Decoder, Meta plans to open the weights of its currently strongest model, Muse Spark 1.2, in the coming weeks. It was released on 5 August 2026 alongside the terminal coding agent Muse Code. The return to an open strategy comes after more than a year during which Meta focused on closed models, and responds, among other things, to enterprise demand for greater control over data and competition from Chinese open models from companies such as Moonshot or Alibaba.
The sources also mention the broader business context: Meta plans to invest up to 145 billion dollars this year and a total of 600 billion dollars by 2028, primarily in data centers, without comparable revenue to match — open models such as Glimmer do not generate licensing fees by design. On the July investor call, Zuckerberg hinted at the possibility of a cloud business without providing specific details, which led to a drop in the share price.
Why it matters
For developers and businesses, Muse Glimmer offers a viable alternative to cloud agents from OpenAI and Anthropic where sending data outside their own infrastructure is a concern — the model fits on a standard graphics card or a Mac. The planned opening of Muse Spark 1.2 weights would also give enterprises access to a stronger model without licensing fees. The essay by Zuckerberg also shows where the dispute over distillation between American and Chinese labs is dividing and how Meta positions open models as its competitive advantage against OpenAI and Anthropic, while investors watch whether a strategy without direct licensing revenue can justify massive infrastructure investment.
Release card
Muse Glimmer
Meta (Meta Superintelligence Labs)
- Specifications
- 30 billion (30B)
- Context
- 128 000+ tokens (128K+)
- Inputs
- Text and images (multimodal); its own 1.8-billion-parameter perception encoder for image input
- Licence
- Apache 2.0
- Availability
- Weights available for download on Hugging Face; running through Ollama (initially only support for Apple Silicon through the MLX engine, with support for NVIDIA, AMD and other platforms expected in the following days)
- Running AI agents locally (tool use, web search, long context) on a single consumer GPU or Mac without the cloud
- Coding and agentic tasks through integrations such as Claude Code, Codex or Pi
- Multimodal tasks involving images – building a website from a design, reading documents and receipts, controlling a computer through screenshots
- Weaker than Qwen3.6-27B at controlling the desktop and terminal, according to the comparison conducted by Meta itself
- Does not compete with the most powerful top-tier closed models – it is a smaller model intended for consumer hardware
- Initial support in Ollama is only for Apple Silicon through the MLX engine; support for NVIDIA, AMD and other platforms is expected in the following days
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
What was added since the original report
Verified updates
-
Mark Zuckerberg published an essay (6000+ words) on AI philosophy, governance and the strategy of Meta; The announcement includes Muse Code, an AI agent for the terminal; Meta strategically positions open-weight models as an alternative to proprietary approaches from OpenAI and Anthropic; The essay criticizes lobbying by competitors and mentions competition from Chinese open-weight models; The new strategy of Meta focuses on differentiation through an open-source approach
- Mark Zuckerberg published an essay (6000+ words) on AI philosophy, governance and the strategy of Meta
- The announcement includes Muse Code, an AI agent for the terminal
- Meta strategically positions open-weight models as an alternative to proprietary approaches from OpenAI and Anthropic
- The essay criticizes lobbying by competitors and mentions competition from Chinese open-weight models
- The new strategy of Meta focuses on differentiation through an open-source approach
-
Meta returns to an open-source strategy after more than a year of focusing on closed models (Muse Spark 1); The release responds to enterprise demand for greater data control and local infrastructure; The model is positioned to compete with the Chinese open-source ecosystem (Moonshot, Alibaba); Glimmer was released in the week after Muse Spark 1.1 (a closed model) – signaling a dual strategy
- Meta returns to an open-source strategy after more than a year of focusing on closed models (Muse Spark 1)
- The release responds to enterprise demand for greater data control and local infrastructure
- The model is positioned to compete with the Chinese open-source ecosystem (Moonshot, Alibaba)
- Glimmer was released in the week after Muse Spark 1.1 (a closed model) – signaling a dual strategy
-
Model available on Hugging Face (not just Ollama); Helper model delivers a 3.1× performance speedup; In the benchmark, it beat Gemma 4-31B and Qwen 3.6-27B on agent tasks; Meta plans an open-weight Muse Spark 1.2; First open model from Meta SI Labs since Llama 4 (2025)
- Model available on Hugging Face (not just Ollama)
- Helper model delivers a 3.1× performance speedup
- In the benchmark, it beat Gemma 4-31B and Qwen 3.6-27B on agent tasks
- Meta plans an open-weight Muse Spark 1.2
- First open model from Meta SI Labs since Llama 4 (2025)
Two audiences, two different impacts
What this means
For individuals
Developers and enthusiasts can run a full-fledged agent for coding or working with files or images directly on their Mac or PC, without depending on the cloud or sending data to a third party.
For a business
Meta offers enterprises an open model as an alternative to cloud services from OpenAI and Anthropic for sensitive agentic workflows, while investors remain uncertain whether the planned investment of 600 billion dollars by 2028 will pay off for the company, because open models do not generate licensing revenue on their own.
StrategyCheck the original
Event sources
confirmed by 4 independent sources · 5 publishers, 4 independent. We count feeds from the same owner only once.