康熙王朝

A |      8 月 26 日消息，Artificial Analysis 于 8 月 24 日发布博文，宣布携手 Liquid AI，发布面向手机端的 AI 跑分基准，重点关注 AI 模型在苹果 iPhone 17 Pro 上的表现。    (ECNS) -- The first phase of the middle route of China's South-to-North Water Diversion Project had transferred more than 80 billion cubic meters of water to northern China as of Friday, directly benefiting nearly 118 million people, according to the Changjiang Water Resources Commission.     The water comes mainly from Danjiangkou Reservoir, where managers balance reservoir inflows, local demand in the Hanjiang River basin, northbound supplies, and downstream ecological needs.        This summer, after forecasts pointed to below-normal rainfall, authorities adjusted reservoir operations to retain about 1.5 billion cubic meters more water as a buffer against possible drought.    Technology is playing a growing role in monitoring water quality.    An AI-powered floating platform works alongside satellites, drones, and ground monitoring stations, feeding data into a digital twin of Danjiangkou Reservoir that tracks water quality, simulates flood scenarios, and supports emergency planning.    At Danjiangkou Reservoir, the core water source for the middle route, an autonomous vessel collects samples from different depths to monitor algae.    Using AI-powered microscopy, the floating station identifies microorganisms quickly and tracks algal activity, giving early warning of blooms.    The vessel is part of an integrated satellite, aerial, surface, and structural monitoring network that combines remote sensing, drones, ground-based stations, and sensors embedded in the dam.    By fusing multi-source data in real time, the system has shifted water monitoring "from isolated points to full-area coverage," said Cui Dongdong, deputy director of the technology development department at a company under the Yangtze River Water Resources Commission of the Ministry of Water Resources.    With the digital platform, a regulation plan that once took days to produce can now be generated in seconds.    The platform runs more than 15 professional models covering flood routing, pollutant dispersion, and dam structure simulation.    The 1,432-kilometer middle route canal carries water across the Yangtze, Huai, Yellow, and Hai river basins before it reaches households in the north.    Water quality has also improved, and the nationally protected "crystal jellyfish," known as the "giant panda of the water," is now sighted frequently in the reservoir, an indicator of excellent conditions.     (By Helen Mo, intern Yang Hongran)                    。其中 Artificial Analysis 负责智能水平测试系统，Liquid AI 负责推理性能测试系统，双方选用 5 项基准测试：BFCL（Berkeley 函数调用排行榜）IFBench（指令遵循测试）AA-Omniscience（知识与认知相关测试）GPQA Diamond（高难度问答测试）MATH-500（数学问题测试）。测试模型对象为 4 bit 量化后体积不超过 8GB 的模型，示例包括 Gemma 4 E2B 和 LFM2-2.6B-Exp。当上下文长度限制为 16K tokens 后，平均基准测试得分最高的是 Nanbeige4.2-3B 和 LFM2.5-2.6B。当响应时间限制为最长 1 分钟时，LFM2.5-8B-A1B 排名第 1，随后是 LFM2-2.6B-Exp、Gemma 4 E4B（Non-reasoning）和 Granite 4.1 8B。Nanbeige4.2-3B 是 BOSS 直聘旗下的南北阁实验室推出，仅含 30 亿非嵌入参数（总参数约 40 亿），采用 Looped Transformer 架构，通过在不增加参数量的前提下复用 Transformer 层（循环两遍），提高模型的有效计算深度和容量。横轴表示“输出 256 个 token 所需的时间（秒）”，纵轴表示“上下文长度限制为 16K 个 token 时的平均基准测试得分”。“Nanbeige4.2-3B”的基准测试得分与“LFM2.5-2.6B”相当。

B | 在上下文长度限制为 16K 个 Token 时，平均基准测试得分中得分最高的是 Nanbeige4.2-3B 和 LFM2.5-2.6B。当响应时间限制为最长 1 分钟时，LFM2.5-8B-A1B 排名第 1，随后是 LFM2-2.6B-Exp、Gemma 4 E4B（Non-reasoning）和 Granite 4.1 8B。附上相关截图如下：。