A roughly 17 GB model file is not a guarantee that every context and tool configuration fits in 17 GB of memory. Muse Glimmer has several components with different roles.
A release and a runtime are separate dependencies
Meta’s August 10 introduction presents a 30B local agent model distilled from Muse Spark. It announced open weights and described several integrations as arriving in the following days, rather than establishing identical day-one support everywhere.
The current official GGUF instructions require llama.cpp build b10353 or newer. They distinguish the text CLI from llama-mtmd-cli for images. The image path needs the perception encoder; attaching that file to an unsuitable text-only invocation is not enough. Check runtime version and follow its current examples before diagnosing the model itself.
Add the components before allowing for context
The published file list gives 16.8 GB for the smaller text build, 1.4 GB for its perception encoder and 1.6 GB for the optional DFlash drafter. Their sum is 19.8 GB of listed files. That arithmetic is not a measured peak RAM/VRAM requirement: runtime buffers, KV cache, placement and other applications remain relevant. The cover separates these layers rather than inventing a successful GPU screenshot.
Context allocation also matters. The instructions explain that a server’s configured context is divided across its parallel slots. In an illustrative 131,072-total, four-slot configuration, each slot gets 32,768 tokens. That is not a 131,072-token budget for every request. Inspect the runtime’s reported slot context, including room needed for reasoning and the answer.
Verify the whole agent loop
A useful local trial includes a successful tool result, an explicit error and an ambiguous timeout. For the timeout case, test whether the agent checks what happened before repeating an operation that may already have succeeded. Good prose and generation speed do not answer that question.
The weights use Apache 2.0, whose redistribution conditions still apply. That license does not certify application reliability or make external agent tools offline. Keep the model’s inference location separate from where its tools send data and perform actions.
September 8: verify official GGUF filenames and runtime floor; separate file sizes from peak memory and text CLI from image CLI; qualify launch support claims.