Meta Muse Glimmer release puts 30B agent model under Apache 2.0
Meta has released Muse Glimmer’s weights under Apache 2.0, targeting local AI agents on high-end consumer hardware.
By Colin Brandt · Enterprise Reporter
· 3 min read
Meta has released the weights for Meta Muse Glimmer, a 30-billion-parameter model under the Apache 2.0 license, with downloads available now. The release gives developers a permissively licensed model aimed at running tool-using agents locally, though Meta’s own memory targets put practical deployment in the high-end consumer hardware tier.
Meta describes Glimmer as an open-weight model for persistent local agent workflows, including function calling, coding and model-based evaluation. Apache 2.0 allows commercial use, modification and redistribution of the released weights. The announcement does not establish that every element of the model’s training stack, data or associated tooling is open source.
For Meta, the licensing change is notable. VentureBeat reported that the company’s April release of Muse Spark was proprietary, making Glimmer its first fully open release since then. Meta CEO Mark Zuckerberg has also said the company plans to release weights for Muse Spark 1.2.
What hardware is needed to run Meta Muse Glimmer locally?
“Local” does not mean an ordinary laptop. Meta says a full-precision 30B model requires more than 55GB of memory. Its approximately 4-bit quantized versions reduce the language-model weights to less than 20GB, with the full operating setup, including the KV cache, image encoder and speculative-decoding component, designed for 24GB or 32GB memory envelopes.
That points to a Mac with substantial unified memory or a PC with a high-end consumer GPU, rather than broadly available 8GB or 16GB systems. Running an agent on the device can reduce dependence on a persistent network connection and keep files, screenshots and other working context off a remote inference service. It also moves compute, electricity, deployment and maintenance costs to the operator instead of eliminating them. For context on the infrastructure being displaced or supplemented, see how cloud infrastructure works.
Meta says Glimmer handles multi-step planning, tool calls and recovery from failed tool calls. It accepts combined text and image inputs through a dedicated perception encoder, supports more than 100 languages, and offers selectable reasoning settings. The model card cited by VentureBeat lists roughly 29.6 billion parameters and a context window of at least 131,072 tokens.
What is available today?
The weights and Meta’s developer documentation are available now. Meta says optimized support for llama.cpp, MLX and ExecuTorch is due in the coming days. Separately, the vLLM project announced day-zero support for Muse Glimmer, allowing developers to serve the model through its runtime.
Meta trained Glimmer by distilling outputs from its larger Muse Spark model, followed by training on longer-context, agent-focused material and post-training across general, reasoning, coding and agent tasks. Meta says its benchmark comparisons show strong results against models in its size class, and says its DFlash speculative-decoding method improves generation speed. Those performance, quantization-quality and speed results are vendor measurements, not independent evaluations.
The release therefore expands the set of locally deployable agent models under a conventional open-source license. It does not settle whether a 30B model will meet a team’s latency, reliability or hardware-cost requirements; those remain deployment questions for developers to test in their own agent scaffolds.
This story draws on original reporting from VentureBeat.
