A 30B model for local agents
Meta released Muse Glimmer 30B on August 10. Its model card describes an approximately 29.6B-parameter dense causal transformer with an approximately 1.8B perception encoder, a context above 131,072 tokens, text-and-image input and text output. The intended capabilities include tool use, multi-step reasoning, failure recovery and end-to-end agent tasks.
The model uses Apache 2.0 and received day-one support in Transformers, llama.cpp and vLLM. Meta says its roughly 4-bit variants can reduce the language model below 20 GB and fit the full setup within 24 GB or 32 GB memory envelopes. Actual memory and speed depend on context length, KV cache, images, runtime and quantization.
Reading the reported benchmarks
The Hugging Face release lists results on MCP Atlas, DeepSearch QA, SWE-Bench Pro and other tests, but these are Meta-reported comparisons. Muse Glimmer leads some rows while Gemma4 or Qwen3.6 leads others, so the table is not evidence of one universal winner. The model card lists text and image input; video demonstrations or third-party integrations should not be conflated with the base supported modality statement.
Deployment boundary
It is a notable candidate for local document analysis, screen understanding and privacy-sensitive workflows, but a 30B model still requires meaningful memory and operational skill. Apache 2.0 does not replace checks on input rights, tool permissions, prompt-injection controls and success rates on the deployer's own tasks.