Meta on Monday released Muse Glimmer, a 30-billion-parameter AI model that runs on a single consumer graphics card, and promised to open the weights of its flagship Muse Spark 1.2 in the coming weeks. The company framed both moves as the opening shots in a push to keep American open-weight models ahead of Chinese rivals.
Chief executive Mark Zuckerberg paired the release with a 6,500-word manifesto titled "The Future is for Everyone: The Path to a Positive AI Future," published on Meta's newsroom. In it he argues that the biggest danger around AI is not the technology itself but too much power concentrated in too few hands, and he calls on Washington to lower barriers for American developers building open models.
"We propose a philosophy based on individual empowerment as the source of prosperity, invention as the primary purpose of superintelligence, and balance of power as the foundation of safety," Zuckerberg wrote.
The timing is deliberate. Chinese open-weight models from Alibaba, DeepSeek, and Moonshot keep picking up users in the United States because they offer competitive performance at a fraction of the cost of closed American systems. Policymakers in Washington are weighing new restrictions on those models over national-security concerns, even as the companies behind them gain ground in enterprises. DeepSeek is already barred from U.S. intelligence-community systems, and lawmakers have floated broader curbs on Chinese open-weight technology.
A model built to live on your desk
Muse Glimmer comes from Meta Superintelligence Labs, the unit Meta stood up as its center of gravity for frontier AI work. At 30 billion parameters it is small next to the frontier systems OpenAI and Anthropic run in the cloud, but that is the point. The Verge describes it as essentially an open-weight version of Muse Spark, Meta's most capable model to date.
The model is engineered for agentic workloads, long-running jobs where the software calls tools, reads screenshots, recovers from failed steps, and works toward a goal over multiple turns. Meta says it handles function calling with precise schemas, chains multi-step reasoning, and retries after a failed tool call instead of stopping dead. A dedicated perception encoder lets it take interleaved text and images, so an agent can interpret charts, documents, and screenshots alongside conversation. The model is trained on data from more than 100 languages, and it supports different reasoning strengths so developers can trade quality against speed.
Everything is tuned to fit on local hardware. At full precision a 30-billion-parameter model needs more than 55 GB of memory, far beyond any consumer GPU. Meta quantizes the weights to roughly 4-bit precision, shrinking the language model to under 20 GB and leaving room for the KV cache, the perception encoder, and a fast draft network inside a 24 GB or 32 GB envelope.
That draft network is the second trick. Muse Glimmer ships with a small companion model based on DFlash that proposes whole blocks of tokens at once; the main model verifies the proposals in parallel. On a MacBook M4-Max, an M5-Max, or an RTX 5090, Meta says the quantized build is fast enough for fluid conversation and real-time agent interaction, entirely on the device. Early third-party numbers agree: AMD reported up to 24 tokens per second on a Ryzen AI Max+ 395 processor and up to 53 tokens per second on a single Radeon AI PRO R9700 card with dFlash enabled, running through llama.cpp with the Vulkan backend.
How it stacks up against rivals
Meta positions Glimmer against the other open models in its weight class. The release notes compare it with Gemma4-31B from Google and Qwen3.6-27B from Alibaba, and Meta says Glimmer performs strongly across agentic benchmarks including DeepSearch QA, MCP-Atlas, and SWE-Bench, which measure end-to-end task completion, tool use, and code repair. The company says it also tested the model under its Advanced AI Scaling Framework and assessed it for open-weight release across all relevant safety categories before shipping.
The use cases Meta lists read like a personal-assistant shopping list: an agent that manages your schedule, drafts your messages, organizes your files, and learns how you work. Those jobs need deep access to personal context, which makes running them on your own machine attractive from a privacy standpoint alone. The model also works across OpenClaw and other agentic orchestration scaffolds, so teams can slot it into existing agent stacks instead of rebuilding around a proprietary runtime.
"Foundation models have achieved remarkable capabilities," Meta's research team wrote in the release post. "Yet most deployments still depend on cloud infrastructure and network access. Running models locally enables you to use AI anywhere, anytime, with or without an internet connection."
The company built Muse Glimmer through a distillation pipeline that transfers agentic reasoning from Muse Spark, its larger teacher model. Pre-training used Spark's outputs with logit distillation, mid-training added longer-context agent-heavy data, and post-training mixed supervised fine-tuning with on-policy distillation and reinforcement learning across reasoning, coding, and agentic domains.
The weights are out under Apache 2.0, the permissive license Meta has used for earlier releases. Hugging Face hosts the model card, and optimized integrations with llama.cpp, MLX, and ExecuTorch are landing over the coming days. Ollama, LM Studio, Unsloth, Together AI, Fireworks AI, and OpenRouter are listed as launch partners, and Meta says developers can fine-tune further with PyTorch's TorchTitan. Hardware partners include AMD, Arm, Dell, Intel, and Nvidia.
The politics of open weights
The release lands at a moment when Washington and Silicon Valley are arguing about how open the most capable models should be. OpenAI recently paused work on Astra after internal tests flagged what it called critical cyber capabilities, and Anthropic has warned publicly about the risks posed by its Claude Mythos line. Zuckerberg is staking out the opposite position.
"I'm personally more worried about centralization than I am about any of the specific risks others are talking about," he told Axios in an interview published alongside the manifesto.
He also warned that American rules are handing an edge to foreign labs. "Foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data," he wrote, adding that even a month-long delay in American model releases would hand momentum to competitors overseas. He argues the United States and its allies should lead the open-source AI ecosystem because it will make up a large share of global AI use.
The manifesto also promises a fully private mode for personal agents, where even Meta or any other service provider cannot see or grant access to user information. That is a direct contrast with Meta AI's current data practices, which let the company train on most interactions with the service.
Anti-doomer messaging is part of the pitch. "I do not understand why anyone who believes that AI will eliminate most jobs and much of humanity's relevance would rush to build that future," Zuckerberg wrote, taking a swipe at the pessimism coming from rival labs.
Meta says its independent board will approve model-release safety criteria and review whether each release meets them. Zuckerberg also promised a "community compact" around data center development, plus a $1 billion Future is for Everyone Fund for U.S. communities where Meta operates data centers, and a skills-trades training program that guarantees jobs at company facilities.
The market gave a muted nod: Meta shares were up about 1 percent before the bell. The real test comes later, when developers decide whether to build on Meta's open models or keep leaning on the increasingly capable Chinese alternatives. The download page went live Monday morning, and the first wave of developers was already pulling the 30-billion weights from Hugging Face within hours. Either way, the frontier of AI competition has moved from the data center to the desk.
Meta's move extends the open-weight debate our AI desk has tracked all week, from OpenAI pausing Astra work over cyber-risk flags to the AI category at large.
For the full technical details, see Meta's release post on Muse Glimmer, the model card on Hugging Face, coverage of the open-weight strategy by NY Post, the Zuckerberg interview with Axios, and AMD's local performance numbers.