Muse Glimmer 30B puts a local agent inside a 24GB memory envelope

The striking part of Muse Glimmer is not simply that Meta released another set of model weights. It is the attempt to fit a 30-billion-parameter agentic model, its working memory, image encoder, and acceleration model into hardware that can sit on a desk. Meta Superintelligence Labs released the weights under Apache 2.0 and positions the model for local coding, function calling, persistent personal agents, and LLM-as-a-judge evaluation.

How the model fits

A 30B model at full precision would require more than 55GB of memory. Meta says it quantized the language-model weights to roughly 4-bit precision, bringing that portion below 20GB. The remaining capacity in a 24GB or 32GB memory envelope can hold the KV cache, the perception encoder used for image input, and a small speculative-decoding drafter. The company reports minimal degradation on its agentic evaluations, though independent testing across hardware and workloads is still needed.

The drafter is based on DFlash. Instead of waiting for the main model to emit one token at a time, the smaller network proposes blocks of tokens and the main model verifies them in parallel, correcting rejected proposals. 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. That matters most during long reasoning traces and repeated tool calls, where latency can make a local agent feel unusable even when its answers are good.

The real test begins after the benchmarks

Muse Glimmer accepts interleaved text and images, can retry after tool-call errors, offers selectable reasoning effort, and was trained across more than 100 languages. The weights are on Hugging Face. Meta also lists planned or emerging support through Ollama, LM Studio, Unsloth, llama.cpp, MLX, vLLM, SGLang, and other runtimes. Those integrations will determine whether the release is a model people can actually operate rather than a benchmark table they briefly admire.

The HN thread, which reached 687 points and 386 comments, quickly moved to hands-on questions. One early user reported acceptable results on a 32GB Mac mini but warned that tasks were slow. Another found the model less prone to repetitive thinking than a comparable local model. Skeptics asked whether public scores would survive private test cases and how it would compare with the next similarly sized Qwen release. Those are early impressions, not settled measurements, but they identify the right criterion: sustained local agent work under real memory and latency constraints.

Meta's technical introduction contains the release details, optimization approach, and model links.