Summary
Sam Witteveen unpacks Meta’s latest releases, led by Muse Spark Open, an 816-billion-parameter mixture-of-experts model with 42 billion active parameters per token, a one-million-token context window, and native video and audio input. Meta calls it the most capable US open-weights model to date. On the Artificial Analysis Intelligence Index it scores 39, behind Chinese open models such as MiMo 2.6 Pro (46), GLM 5.3 (45), Kimi K3 (44), and Qwen 3.8 (40).
The video compares it with Reflection’s Beam, a 501-billion-parameter model with 23 billion active that scores 30 on the same index. Spark Open wins most published benchmarks, including MCP Atlas (89 vs 79) and DeepSuite (60 vs 44), while Beam leads on Terminal Bench (80 vs 76). Beam is smaller, more token-efficient, and Apache 2 licensed, whereas Spark Open uses Meta’s Muse license, which requires attribution and a separate agreement for large model-as-a-service businesses.
Witteveen also clarifies that Meta’s popular Muse personal agent runs on a customized, non-public model, not Spark Open. The video describes how Meta built it: a spreadsheet of more than 100 target behaviors, each with its own evals and launch thresholds, a team under 200 people, and about seven months of work, offering a template for building agentic products.
📺 Source: Sam Witteveen · Published October 19, 2026
🏷️ Format: News Analysis
