On August 10, 2026, Meta Superintelligence Labs introduced Muse Glimmer, a 30-billion-parameter agentic model released as open weights on Hugging Face under an Apache 2.0 license. The positioning is unusually specific: this is a model built for always-on local agent workflows, small enough to run on a Mac or a PC with a single consumer GPU. The intended use cases are local agents and function calling, local coding, and LLM-as-a-judge evaluation.
It is also Meta’s first open-weights release in more than a year. After Llama 4 flopped, the company restructured its AI group, and critics openly questioned Mark Zuckerberg’s commitment to open source AI, this launch carries more weight than “yet another small model” normally would.
30B Parameters, One Consumer GPU
At full precision, a 30-billion-parameter model would need over 55 GB of memory — more than any consumer GPU offers. Meta’s answer is quantization. The Register’s analysis puts the 4-bit version at just under 16 GB, small enough to fit entirely on 20-to-24 GB cards like the RTX 3090 or 4090 and the RX 7900 XT/XTX. Owners of 16 GB cards would have to drop to a less accurate 3-bit quant.
The speed figures are respectable for local hardware. Meta says an RTX 5090, with its 1.8 TB/s of memory bandwidth, can push 75 to 233 tokens per second, with the top end relying on speculative decoding using a small draft model. On an M5 Max MacBook Pro, the estimate is a still-usable 26 to 58 tokens per second. On the software side, optimized integrations for llama.cpp, MLX, and ExecuTorch were promised within days of the announcement, and Ollama and LM Studio already offer downloads.
Simon Willison’s early hands-on highlighted two things. First, the license: he called Apache 2.0 “a step up from the janky Llama licenses of old,” which matters for anyone embedding the model in a product. Second, the model is multimodal — he had it describe a photograph in detail and ran his coding-agent plugin against a fresh checkout of Datasette, watching it explore the codebase with tool calls to answer a question about how authentication works. His practical takeaway: on machines with 32 GB of RAM or more, a model this size leaves plenty of room for everything else you are running.
Distilling Agentic Capability from a Teacher
The announcement is candid about what makes a local agent hard. An agent that manages your schedule or organizes your files needs long-horizon execution, precise tool calling, multimodal understanding, long-context memory, and instruction following — all at once, and all inside the memory and compute budget of consumer hardware. Meta’s approach is a novel distillation recipe that transfers agentic reasoning from a much larger teacher model. The Register identifies the teacher as Muse Spark, Meta’s proprietary flagship.
On benchmarks, Meta’s own numbers compare Muse Glimmer against Gemma4-31B and Qwen3.6-27B, the obvious peers in its size class. The evaluations Willison quoted go further, covering full-task benchmarks — DeepSearchQA, MCP-Atlas, τ-Bench, and SWE-Bench — which measure whether a model can actually work within scaffolds, write and debug code, and carry multi-turn requests to completion rather than just answer trivia.
Why Meta Returned to Open Weights
The timing is not subtle. The Register frames it bluntly: Chinese open-weights models — Kimi K3, Qwen 3.8-Max, DeepSeek V4 Flash — have dominated the AI news cycle for months, and American companies are wrestling with the lack of credible domestic open alternatives. Meta built its reputation on the Llama family, then went quiet after Llama 4 disappointed. Muse Glimmer ends that silence.
There was a second announcement hiding inside the same day: Alexandr Wang, Meta’s superintelligence chief, posted on X that an open-weights version of Muse Spark 1.2 is coming “soon.” Muse Spark 1.2 is Meta’s most capable model, its first to meaningfully compete with OpenAI, Anthropic, and Google. Read together, the message is that Glimmer is the advance party and the flagship’s open version is the main course. Whether that release keeps pace with the Chinese frontier is the real test of Meta’s renewed open-source commitment.
What It Means for the Local AI Ecosystem
Three things are worth tracking. First, an open 30B model with genuine agentic training raises the floor for local coding assistants and personal agents — agents that run without network access or API keys move from demo to daily driver. Second, Apache 2.0 removes the licensing friction that made enterprises hesitate around the old Llama terms, and community fine-tuning groups such as Nous Research will almost certainly ship derivatives. Third, keep the limits in view: as a dense model, Glimmer spends all 30 billion parameters on every token, so it is inherently slower than sparse designs like Qwen 3.6-35B-A3B. And The Register notes the benchmark comparison may not age well, with Qwen’s next 27B model expected imminently.
Sources
- Introducing Muse Glimmer — Meta AI Research
- Zuck rekindles open weights Llama drama with Muse Glimmer — The Register
- Introducing Muse Glimmer — Simon Willison
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.