Meta Fits 30B Muse Glimmer Under 20GB for Local AI Agents

A 30B agent model moves onto consumer hardware

The hardware threshold for an always-on personal agent has dropped. Meta's Muse Glimmer retains 30 billion parameters while compressing its language-model weights below 20GB, targeting Macs and PCs equipped with a single capable consumer GPU. That matters because agents handling schedules, messages, and files need deep access to personal context. Local execution can keep more of that context on the user's machine while removing a continuous dependency on network access and cloud inference.

“Introducing Muse Glimmer, an open-weight 30B-parameter model optimized for local, always-on agent workflows.” — AI at Meta

The source post says the weights are available under the permissive Apache 2.0 license. AI at Meta is the company's official research account, normally used to publish research models, datasets, evaluations, and links to technical material. This release points readers to a detailed research post, a Hugging Face model repository, and developer documentation rather than relying on the social post alone.

Four-bit quantization and DFlash expand the deployment envelope

At full precision, a 30B model would require more than 55GB of memory. Meta applies roughly four-bit quantization to bring the language model below 20GB. The remaining headroom is intended for the KV cache used during long contexts, a perception encoder for image input, and a speculative-decoding drafter. Meta targets complete deployments within a 24GB or 32GB memory envelope and says its internal agent evaluations found minimal to no degradation from compression. Independent reproduction will be important because that quality claim comes from the model maker.

Generation speed is improved with a lightweight drafter based on DFlash. It proposes blocks of tokens that the main model verifies in parallel, instead of producing every token through the full model sequentially. Meta reports a 3.1× decode-speed increase on an RTX 5090, 1.8× on an M5 Max, and 1.5× on an M4 Max. Lower latency is especially important for agents that must reason, call a tool, inspect the result, recover from errors, and continue across many steps.

Benchmarks, licensing, and the next test

Meta compares Muse Glimmer with Gemma4-31B and Qwen3.6-27B across agentic, coding, multimodal, safety, and reasoning evaluations. The listed suites include DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench. The model accepts interleaved text and images, was trained on more than 100 languages, and is designed to work with agent scaffolds such as OpenClaw. The central claim is not a single overall leaderboard victory; it is the combination of competitive size-class performance, tool use, and practical local deployment.

The weights are available from the Muse Glimmer repository. Optimized llama.cpp, MLX, and ExecuTorch integrations are due in the coming days, with support from Ollama, LM Studio, and Unsloth also planned. The next things to watch are real token throughput on 24GB systems, reliability during long tool-calling runs, independent replication of Meta's benchmark table, and whether Apache 2.0 distribution accelerates a broader ecosystem of private local agents.