Meta open-weights Muse Glimmer, a 30B agent model that runs on a single consumer GPU
Meta open-weighted Muse Glimmer today. 30 billion parameters, Apache 2.0, designed for local agentic workflows. With 4-bit quantization it fits in 18-20 GB of VRAM, which puts it on a single consumer GPU — a 24 GB RTX 5090, a 32 GB Mac, an AMD Ryzen AI Max box. It runs at over 20 tokens per second on Blackwell Ultra and around 24 on the AMD hardware Meta partnered with. Muse Spark 1.2, the closed model it is distilled from, is promised as open weights in the coming weeks.
The benchmarks land where the framing suggests they should. On GAIA2 it scores 43.3 against Gemma4-31B's 36.4 and Qwen3.6-27B's 40.0. On SWE-Bench Pro it takes 51.2 against Gemma's 36.9 and edges Qwen's 50.2. On SWE-Bench Verified it trails Qwen at 76.0 to 77.2. Terminal-Bench 2.1 is where it loses to Qwen3.6-27B, 51.7 to 60.7. Read those numbers together and you have a competitive local agent model, not a flagship-killer. That is exactly the tier that reshapes build-versus-buy for verticals.
I have been running through the routing table for two client stacks this week, and here is what changed. For coding-adjacent agents, the deployment target used to be a hosted flagship or a self-hosted 70B on a rented A100 box. Now the same coding-agent scaffold — write, run, patch, verify — has a plausible local backend that a small team can run on hardware they already own. The cost math tightens. The privacy math tightens more.
The distillation lineage matters. Muse Glimmer is distilled from Muse Spark, the closed model Meta shipped last month. Zuckerberg's framing is that personal superintelligence should be something you own, not something you rent. Push on the framing if you want; the models are the datum.
For anyone building agents, three concrete things to test this week.
One, the does-it-work-in-my-scaffold test. GAIA and SWE-Bench Pro are useful, but the number that matters is task completion in your own eval harness. The 37% gap between benchmark scores and real deployment performance that keeps showing up in agent studies is not a benchmark problem; it is a scaffolding-and-tools problem, and it shows up sharpest on local models. Run your top-50 real user queries. That is your eval.
Two, the LLM-as-a-judge test. Meta calls out judge use as a first-class case. That is the workload where the economics of a local model swing the hardest — you are running it on every trace, every session, every synthetic eval. If your judge budget has been a line item this year, this is the model to swap in and re-run.
Three, the recovery-model test. This is the lens I picked up from the July Hugging Face post-mortem. When the frontier models refused to help incident responders because their guardrails could not distinguish a defender from an attacker, teams reached for open-weight models. Muse Glimmer is a plausible answer there — self-hosted, small enough to boot on hardware in your incident-response bunker. Do not wait until you need it to find out whether it works.
What I am not sold on. Terminal-Bench trails Qwen3.6-27B by nine points. If your agent lives in the shell — the compliance auditor, the SRE, the DevOps bot — Qwen may still be the better local backend for now. Do not switch on brand.
What I am watching for. Alibaba is scheduled to open-weight Qwen3.8-Max and Qwen3.8-27B this same week. If those land in the next few days, we get a genuinely competitive local three-way. The build-versus-buy question in narrow verticals — legal extraction, medical coding, compliance drafting — flips from "which hosted API" to "which open-weight base do I fine-tune."
One last note on framing. The gap between "runs on a laptop" and "runs in production on a laptop" is still wide, and I would not confuse the two. Muse Glimmer is not going to sit under a customer-facing chatbot at 500 QPS on a consumer GPU. It is going to sit under a developer tool, an internal agent, an eval judge, a research assistant. That is a large market. It is not the same market as the hosted flagships.
If you have a vertical AI product and you have been paying frontier prices for the workhorse tier, reply — I am collecting the routing tables people redraw this week.