Meta has released Muse Spark, a new proprietary LLM developed by Meta Super Intelligence Labs over the past nine months, likely the model previously codenamed 'Avocado'. Unlike the Llama series, it is closed-source with no public API yet. The model comes in instant and thinking modes, supports multi-agent capabilities, and integrates with services like Google Calendar and Gmail. Benchmarks show it competitive but not leading — sitting in the top five per third-party analysis, behind GPT, Gemini, and Claude Opus, and last on some benchmarks including Humanity's Last Exam. The release raises questions about whether Meta will open-source it, and whether the massive investment in talent and acquisitions (including Manis) has paid off. The model appears well-suited for Meta's consumer platforms like Facebook and WhatsApp, but its agentic performance is underwhelming given expectations.
Sort: