Meta reclaims leadership in US open-weight models with Muse Spark Open

Summary

Meta's Muse Spark Open model, featuring 816 billion total and 42 billion active parameters and a 1 million token context window, has reclaimed the lead among US open-weight models. It has nearly halved the performance gap with top Chinese frontier models. In this process, Meta is also deploying its closed models (Muse Wonder) to challenge Anthropic and OpenAI.

Why it matters

The US open-weight ecosystem was in a visible crisis after more than 20 Chinese models surpassed the best Western versions. The emergence of Muse Spark Open, alongside Reflection's Beam and Mistral Large 4, signals a competitive turning point that could reshape inference providers and enterprise adoption. This is critically important as Meta is the only American lab releasing both frontier open and closed models. The company's strategic positioning and licensing choices serve as a leading indicator for the entire Western AI sector.

Highlights

• The model is likely the security-tuned Muse Spark 1.2 version, not 1.3. There is only a slight drop of about 1 point on the Artificial Analysis index to reinforce safety behaviors.
• The proprietary license mandates 'Developed with Muse' branding, imposes a $20 million monthly revenue threshold, and requires an acceptable use policy. These conditions move the model far from being truly open source according to OSI standards.
• Meta's unusually flat organizational structure enables a consistent and rapid release cadence. Alex Wang having over 200 direct reports allows this speed to compete with or outpace more bureaucratic labs.

Critical take

The 'US open weights are back!' framing slightly overstates the recovery. Muse Spark Open is still behind GLM 5.3 and Kimi K3. Furthermore, restrictive license terms such as branding requirements, revenue thresholds, and acceptable use policies weaken the 'open' narrative the text relies on. The author also shows significant hesitation about whether Meta is at the first or second tier, indicating that the competitive landscape is less clear than the headline implies.