NVIDIA has just claimed the top spot in U.S. open-source AI models, only for Meta to overtake it with a score of 39, while its agent benchmark lags behind Chinese models.

Beating AI News Flash from Dong察: NVIDIA’s U.S.-based open-source model held the top spot for just a few hours. NVIDIA’s Nemotron 3.5 Super VL scored 26.5 points on the Artificial Analysis Intelligence Index, outperforming all previous U.S. open-source models. But Meta’s subsequently released Muse Spark Open scored 39 points, 12.5 points ahead of NVIDIA, to become the new U.S. leader. However, when looking at the global open-source model rankings, Meta still lags behind multiple Chinese models. Xiaomi’s MiMo-V2.6-Pro leads with 46 points, Zhipu AI’s GLM-5.3 scores 45 points, while Muse Spark Open ties with DeepSeek V4.1 Flash at 39 points. Muse Spark Open is already close to Meta’s closed-source flagship Muse Spark 1.2, launched in August, which scored 40 points. Among the top 9 highest-scoring open-source models, Muse Spark Open ranks last in the combined performance of 4 Agent tests, but places fourth in the total of the other 6 tests. The 6-point gap between it and Zhipu AI’s GLM-5.3 comes entirely from the Agent tests. On the contrary, it leads all open-source models with 25 points in the knowledge accuracy test AA-Omniscience, with a hallucination rate of only 24% – the lowest among open-source models scoring over 30 points. Muse Spark Open also has strong token utilization efficiency: it outputs an average of about 40,000 tokens per test, 43% less than GLM-5.3. The two have similar API unit prices, but because Muse consumes fewer tokens, its average cost per test is only $0.84, less than half of GLM-5.3’s, though still about 3 times that of DeepSeek V4.1 Flash.