Meta's Muse Glimmer 30B excels at tool use but struggles with control and safety

Meta has released two new models, Muse Spark 1.2 and Muse Glimmer 30B, with Glimmer being an open-weights model distilled from Spark. While Spark 1.1 (an earlier version of Spark) leads the MCP-Atlas leaderboard for tool-use protocols, Glimmer is not yet on this benchmark. Meta's own figures show Glimmer excelling at protocol-based agentic tasks but underperforming in terminal and desktop control compared to models like Qwen3.6-27B. Additionally, Glimmer exhibits a notable vulnerability to prompt injection attacks, with a 28.4% attack success rate. AI

IMPACT New open-weight models with specialized capabilities could accelerate agent development, but safety concerns require careful consideration.

RANK_REASON The item details new models and their performance on specific benchmarks, including an analysis of their strengths and weaknesses. [lever_c_demoted from research: ic=1 ai=1.0]

Read on dev.to — MCP tag →

• Apache 2.0
• Claude Fable 5
• Claude Opus 5
• Gemma4-31B
• MCP-Atlas
• Meta
• Muse Spark 1.1
• Muse Spark 1.2
• OpenAI
• OpenRouter
• Qwen3.6-27B
• Siren AgentDojo

• model release
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →