1、GLM-5.3在保持7440億參數規模不變的情況下，主要通過後訓練擴展提升性能，在編碼、長任務執行、成本效率和網絡安全等基準上顯著進步，並據稱在多項測試中超過1.6萬億參數的DeepSeek V4 Pro。 2、 智譜的核心護城河不只是模型，ZCode活躍用戶已超過100萬，形成了“用戶使用→真實任務數據→反哺模型訓練”的數據飛輪，有助於維持GLM在代碼場景中的性價比優勢。閒時任務機制有助提高推理基礎設施利用率。 3、行業定價出現正常化跡象，智譜編碼套餐折扣收窄，DeepSeek V4 Pro API價格相對V4 Flash大幅提高，顯示前沿模型能力提升正轉化爲一定定價權。 4、高盛維持智譜“中性”評級，目標價1610港元。其認爲智譜長期商業化空間仍大，當前ARR可能低估真實需求，未來還可通過開放權重和商業授權釋放收入潛力。 5、風險在於其估值仍高，且面臨競爭加劇、高研發投入、持續虧損和地緣政治風險。（上述觀點來自高盛8月17日報告）

U.S. 6-week Treasury auction settling Aug. 18 awarded yield 3.645%, vs 3.67% previous.

Three sources said on Monday the US and Canada are negotiating a possible cut to US auto tariffs from 25% to 15% after deducting some value content. If no deal is reached by a Wednesday deadline, the US will impose heavy new tariffs on US$20 billion of other Canadian goods. US officials propose the 15% tariff would exclude only US domestic value; Canadian officials want to exclude all North American content, including parts from Canada and Mexico. Auto industry officials said a broader regional-content deduction could cut the effective tariff on North American-made cars to the low single digits, well below the 15% applied to imports from Japan, Korea and the EU and the 10% tariff on most UK cars.