Meta Superintelligence Labs has released Muse Glimmer, a 30-billion-parameter model built for agents that run on a user’s own machine, and it put the weights on Hugging Face under an Apache 2.0 license. Developers can download, modify and build products on it without asking Meta’s permission.
Meta’s research blog says the model is small enough to run on a Mac or PC with a single consumer GPU, with or without an internet connection. It was distilled from Meta’s larger, closed Muse Spark model, trained on Spark’s outputs and then on longer-context, agent-heavy data.
What it takes to run
At full precision, Meta says a 30-billion-parameter model would need more than 55GB of memory. Quantizing the weights to roughly 4-bit precision brings the language model under 20GB, which leaves room for the KV cache, the image encoder and a speculative decoding drafter inside a 24GB or 32GB envelope. Meta said the compression causes “minimal to no degradation on agentic tasks.”
Digital Trends broke down the two quantized builds: K-Quant-Dynamic needs 32GB and loses about 0.2% accuracy, while K-Quant-17GB fits in 24GB with about a 1% hit. In practice that means an RTX 5090, 4090 or 3090, or a Mac with an Apple silicon Max chip. Meta tested on the M4 Max, M5 Max and RTX 5090, and Mashable noted that the average work laptop won’t make the cut.
Speed comes from a lightweight drafter based on DFlash, which proposes blocks of tokens that the main model then checks in parallel. On an RTX 5090, Digital Trends reported, that lifts output from 74.9 tokens per second to 233.4.
Built for agent work
Meta says it trained Glimmer for function calling with precise schemas, long multi-step plans and recovering from failed tool calls by diagnosing the error and retrying. It accepts interleaved text and images through a dedicated perception encoder, answers in text, was trained on data from more than 100 languages and works with OpenClaw and other agent scaffolds. Digital Trends put its context window past 131,000 tokens and its knowledge cutoff at Jan. 4, 2026.
On benchmarks including DeepSearch QA, MCP-Atlas, τ-Bench and SWE-Bench, Meta compared it with Google’s Gemma4-31B and Alibaba’s Qwen3.6-27B. Digital Trends said Glimmer leads those two on planning and multi-step tasks but trails when it has to operate a desktop directly. Mashable pointed out that Meta’s results haven’t been independently verified.
Meta said local runs through Ollama, LM Studio and Unsloth, edge deployment through llama.cpp, ExecuTorch and MLX, and serving through vLLM and SGLang would follow in the coming days. It is also working with AMD, Arm, Dell, Intel and NVIDIA on device-specific tuning.
Open weights, not open data
Mashable noted that “open weight” is the more precise label than open source: the trained weights are public, but the training data and infrastructure aren’t. CNET compared the model’s size with Moonshot AI’s Kimi K3, which has 2.8 trillion parameters.
Mark Zuckerberg tied the release to an essay arguing that access to advanced AI shouldn’t be limited to a few institutions. “The defining questions of our age are who will have access to superintelligence and what will we direct it towards,” he wrote. He also said Meta will soon release the weights for Muse Spark 1.2, the larger model Glimmer was distilled from.
WebProNews is an iEntry Publication