Zhipu GLM-5.2 corona el primer índice inteligente AA de código abierto: GDPval iguala en puntuación a GPT-5.5

robot
Generación de resúmenes en curso
Golden Finance reports that the latest MoE flagship model GLM-5.2 by Zhipu AI scored 51 points in the Artificial Analysis large model intelligence index v4.1 evaluation, surpassing MiniMax-M3 (44 points), DeepSeek V4 Pro (max, 44 points), and Kimi K2.6 (43 points), topping the global open-source model leaderboard.
In the GDPval-AA v2 test simulating real-world knowledge work, GLM-5.2 scored 1524 points (human benchmark 1000 points), leading MiniMax-M3 (1418 points) and DeepSeek V4 Pro (max, 1328 points), and tying with the closed-source cutting-edge large model GPT-5.5 (xhigh reasoning). Compared to the previous generation GLM-5.1, scientific reasoning CritPt improved by 16 percentage points to 21%, HLE increased by 12 percentage points to 40%, TerminalBench v2.1 rose by 16 percentage points to 78%, and GPQA Diamond reached 89%.
GLM-5.2 occupies the best cost-performance position on the "Intelligent - Task Cost" Pareto frontier. Since the average output per task is 43k tokens (compared to 26k for GLM-5.1), the average cost per single task for GLM-5.2 has risen to about $0.46, higher than GLM-5.1 ($0.25) and DeepSeek V4 Pro (max, $0.05), but still far below other models in the same intelligence tier.
GLM-5.2 has a total of 744B parameters, with 40B active parameters, and the context window has increased from 200K to 1M compared to its predecessor, following the MIT open-source license. Currently, Zhipu's official API (pricing input 1.4, output 4.4 / per million tokens) is available on platforms such as SiliconFlow, DeepInfra, Nebius AI, and others.
Ver original
Esta página puede contener contenido de terceros, que se proporciona únicamente con fines informativos (sin garantías ni declaraciones) y no debe considerarse como un respaldo por parte de Gate a las opiniones expresadas ni como asesoramiento financiero o profesional. Consulte el Descargo de responsabilidad para obtener más detalles.
  • Recompensa
  • Comentar
  • Republicar
  • Compartir
Comentar
Añadir un comentario
Añadir un comentario
Sin comentarios
  • Fijado