Meta launches Muse family for agentic workflows

Meta has rolled out a new family of AI models under the Muse banner, positioning Wonder as its flagship, Spark as a cost‑effective middle tier, and Ember for high‑volume, repeatable agent tasks. The shift replaces the Llama lineup and adds open‑weight options like Glimmer for on‑device agents.

What happened

According to the Zapier blog, Meta’s Muse lineup now includes:

• Muse Wonder – the most capable model, offered via API and in Muse Code, with a “Full” variant available only through a Muse Code subscription. It introduces Active Memory (state persistence across conversations) and Active Focus (goal‑oriented error recovery). Pricing: $4 / M input tokens, $20 / M output tokens.
• Muse Spark 1.3 – a multimodal, tool‑using model that supports text, images, audio, and video. It is cheaper and faster than Wonder, costing $1.25 / M input tokens and $4.25 / M output tokens. A “Contributor” tier drops the price to $0.10 / M input and $0.20 / M output in exchange for data sharing.
• Muse Ember – a lightweight model built for repetitive agent work (tool calls, routing, extraction). It has a 1 M‑token context window, can emit up to 512 k tokens, and costs $0.10 / M input and $0.50 / M output.
• Muse Glimmer – an open‑weight 30 B‑parameter model designed to run locally on a consumer GPU, downloadable from Hugging Face.
• Muse (personal assistant app) – uses Spark under the hood, runs each user on a dedicated “Muse Secure VM” Linux instance, and includes a Sentinel watchdog for safety.
• Muse Code – a terminal‑based coding agent that can run any of the Muse models. When paired with Wonder Full, its performance matches top‑tier offerings from Claude and OpenAI.

Developers can access Spark, Wonder, and Ember through the Meta Model API, and all three can be wired into Zapier automations that trigger across 9,000+ apps. Meta also announced a voice‑first device, Muse Charm, though details are sparse.

Why it matters for small teams

For builders of small‑to‑mid‑size companies, the Muse family offers a clear scaling ladder. Cost is a primary concern: Ember’s sub‑$0.20‑per‑million‑token rates make it viable for batch‑style processing (e.g., document parsing or routing), while Spark provides a middle ground for multimodal tasks without the premium price of Wonder. Agentic features like Active Memory and Active Focus reduce the need for custom state‑management code, letting teams focus on business logic rather than prompt engineering. The open‑weight Glimmer model lets privacy‑sensitive teams run agents on‑premises, avoiding data‑exfiltration worries while still supporting the same tool‑use patterns. Finally, tight integration with Zapier means teams can embed Muse models into existing no‑code workflows, gaining observability and governance without building a full orchestration stack. In short, Meta’s Muse lineup gives small teams a tiered, price‑transparent way to experiment with and graduate to production‑grade AI agents.

What to do this week

1. Set up a Zapier automation that calls Muse Spark for a repetitive task (e.g., extracting leads from emails) and measure token usage against your budget.
2. Download Muse Glimmer from Hugging Face and run a simple local agent on a GPU‑enabled workstation to test latency and data‑privacy guarantees.
3. Compare the token‑costs of Ember versus Spark for a batch‑processing job and decide whether upgrading to Wonder is justified for any high‑complexity use case.

Sources

• Muse Wonder, Muse Spark, and Muse Ember: A guide to Meta's AI models — Zapier Blog