Meta released Muse Glimmer on 10 August 2026, a 30-billion-parameter agentic AI model distilled from its larger Muse Spark family and licensed under Apache 2.0 — genuinely open weights, no usage restrictions, downloadable from Hugging Face. The headline feature is where it runs: Muse Glimmer fits on a single consumer GPU with around 24GB of VRAM, the kind of card a well-equipped developer workstation already has, rather than requiring cloud API access at all.
This lands five days after Meta launched Muse Code, its paid, hosted terminal coding agent, and the contrast is the actual story. One product is a cloud-metered competitor to Claude Code and Codex; the other is a free-to-download, run-it-yourself model aimed at developers and companies that want an agent with no per-token bill and no data leaving their own hardware. Meta hasn’t picked a lane — it’s running both bets simultaneously, and Zuckerberg and AI chief Alexandr Wang have said open weights for the larger Muse Spark 1.2 model are coming too.
So what
For teams evaluating AI coding tools, “which vendor” is no longer the only axis worth thinking on — “hosted API versus self-hosted open-weight model” is now a genuine trade-off among frontier-adjacent options, not just a budget compromise. A locally-run agent removes per-token cost and data-residency concerns that matter for regulated or security-conscious clients, at the cost of the infrastructure and tuning work to run it well. We help teams pick and integrate the right AI development stack for what they’re actually building, not just what’s cheapest on paper — see our AI-assisted development approach or get in touch to talk through your options.