Meta releases closed-weight Muse Ember for agent tasks, fueling open-source retreat debate

Meta has added a small, cheap model to its Muse family. Muse Ember 1.0 is designed for agent execution and high-volume repetitive tasks, the kind of work where per-token cost matters more than peak reasoning. It's available as a hosted model, and Meta's Muse Ember, Muse Spark 1.3 and Muse Wonder models are now offered through Google Cloud, giving Meta a route into enterprise buyers who already run on Google's platform.

The biggest detail is what Meta didn't release. Muse Ember comes with no published weights or license terms, a clear break from the open-weight Llama approach that made Meta the default open model provider for much of the industry. Combined with reports that DigitalOcean dropped Llama 4 Maverick, the community is debating whether Meta is retreating from open weights altogether. Llama hasn't disappeared, though: it still powers internal experiments such as the Arena prediction-market app.

Meta also expanded Muse's consumer agent features. US paid subscribers can now have Muse make outbound phone calls to businesses, putting Meta in competition with Google's earlier calling-agent work and with OpenAI's push into agentic commerce through Business Agents and ChatGPT Wallet.

Competitively, Ember enters a crowded cheap-model market: Google's Gemini 4 Hydrogen at $1.50/$7.50, DeepSeek's deepseek-flash, xAI's Grok 4.7 and Mistral's open-weight Large 4 all target similar workloads. Without weights, Ember loses the self-hosting and sovereignty argument that Mistral is now actively making. Watch for published pricing, independent benchmarks, and whether Meta clarifies its open-weight plans for future Llama releases. For developers who built on Llama, that's the bigger question.