Meta releases Muse Glimmer, a 30B Apache-2.0 agent model

Meta Superintelligence Labs has put out a 30-billion-parameter agentic model called Muse Glimmer under an Apache 2.0 license, with weights on Hugging Face and day-zero support in the usual local-inference stacks. The developer page frames it as an always-on model for a Mac or PC with a single consumer GPU, and the Hugging Face blog post spells out the architecture: a 2B ViT-style Perception Encoder bolted onto a 28B parameter text decoder, distilled from a larger internal model called Muse.

The reason to pay attention is not the raw benchmarks, though Meta's own tables put Glimmer ahead of Gemma4-31B and Qwen3.6-27B on MCP Atlas, GAIA2, SWE-Bench Pro and AIME 2026. It is the packaging. A permissive open license, a 30B footprint that fits under 20 GB after 4-bit quantization, tool calling and multimodal input trained in from the start, and a launch-day matrix of runtimes: llama.cpp, MLX, ExecuTorch, vLLM, SGLang, Ollama and LM Studio, plus hosted access through Together AI, Fireworks AI and OpenRouter. Phoronix reports that Ollama picked up support in version 0.32.7 the same day.

Local agentic workloads are the specific bet. The Hugging Face writeup describes the model as "especially designed for local agentic use cases," listing coding, document analysis and personal assistants as the intended surface, with training aimed at end-to-end task completion, precise tool calls, multi-step reasoning, and recovery when a tool fails. Full precision reportedly needs more than 55 GB of memory; the 4-bit K-Quant build brings decode speedups Meta lists as 3.1x on an RTX 5090 and 1.8x on an M5 Max, which is the shape you want if the pitch is a laptop-class always-on agent.

The obvious caveat is that the benchmark numbers are self-reported and the comparison points are Meta's own choices, so a Gemma or Qwen team could pick different tests and get a different story. The developer page and the Hugging Face post also say very little about training data provenance or whether a separately instruction-tuned or safety-tuned variant is coming, and the flashy speedup numbers are quoted at 4-bit, not the BF16 setting used for the benchmark tables.

If those numbers hold up in independent testing, the downstream move is on the pricing side. Anyone paying per-token for hosted tool-calling on evaluation or coding-agent workloads suddenly has a self-hostable 30B alternative from the same company that ships Llama, and the small teams and privacy-sensitive users priced out of hosted frontier access get a real option.