Meta shakes up the frontiers, open and closed

This week marked a significant shift in the American open-weight model landscape towards the frontier. Meta's Muse Spark Open has emerged as the strongest American-made open model by a considerable margin, while Reflection's first model, Beam, also adds relevance. The absence of these two models would see 20+ open-weight models from Chinese companies outperforming the best Western models.

However, despite Mistral's announcement of a strong model, Mistral Large 4, with an index score of 38, its weights are yet to be released. Rumors suggest more strong American open-weight models might follow soon.

The gap between the best American and Chinese models has narrowed approximately by half. Previously, the best American open models were largely irrelevant to consumers needing cutting-edge performance, but now, they are merely behind but still relevant. Meta's release of Muse Spark Open's weights before the others has set the stage for a competitive frontier model business. The industry is now questioning the absence of Kimi and Zhipu AI's recent releases.

While Muse Spark Open's adoption seems promising, its success remains uncertain. The open ecosystem has witnessed a surge in demand following the rapid performance increase of Chinese models. Inference providers are scaling rapidly, the open post-training economy is forming, and numerous enterprises are seeking to own their intelligence.

Muse Spark Open, with its 816B total/42B active parameters and 1M token context window, is designed for coding and agentic work, making it multimodal and safety-post-trained for robustness against misuse. It is positioned somewhere between GLM 5.2, DeepSeek V4.1 Flash, and ahead of GLM 5.3 or Kimi K3, making it "good enough" for enterprise adoption.

However, the final outcome of Muse Spark Open's success will depend on various factors, which will be explored further in a subsequent piece.