TSMC reports August revenue of NT$514.8 billion, a 53.3% year-on-year increase.

Beating AI Express: Over the past few months, the top spot among domestic Chinese AI models has been swapped between Kimi and Zhipu AI. In the latest third-party ranking from Artificial Analysis, Kimi K3 and GLM-5.3 scored 44 and 45 points respectively, while DeepSeek V4 Pro 0813 only earned 36 points. Now DeepSeek is fighting back with its V4.1 Flash model. According to official evaluations, among the 16 tests where both models have results, V4.1 Flash won 13 against Kimi K3; in the 13 overlapping tests with GLM-5.3, it took 11 wins. Its advantages lie mainly in coding, terminal operations, and agent automation. What’s more, this may just be the start: Kimi K3 is Kimi’s current strongest 2.8T parameter model, GLM-5.3 is Zhipu’s flagship, while V4.1 Flash is merely the smallest model in DeepSeek’s new architecture series.

According to Bitget market data, South Korea’s KOSPI index closed 16.77 points lower today, down 0.24% to 7034.87 points. SK Hynix gained 0.43%, and Samsung Electronics rose 0.09%. Japan’s Nikkei 225 index closed 128.17 points higher today, up 0.20% to 65270.95 points. Kioxia climbed 2.07%, while SoftBank edged down 0.26%.

According to an official announcement, Huobi HTX launched perpetual contracts for USDJPY/USDT, FCX/USDT, XLK/USDT, and GE/USDT on September 10, supporting long and short positions with leverage ranging from 1x to 20x. Meanwhile, from now until 15:00 UTC+8 on September 15, Huobi HTX is hosting a new contract token trading contest. Users who complete registration, trade the specified contracts, and meet the required thresholds will have the chance to split a total prize pool of 1 billion HTX tokens.

According to Bitget market data, Hong Kong-listed software application stocks have continued to plunge sharply. Qunhe Technology, Zhipu AI, and Unisound all fell more than 10%; MiniMax dropped over 9%; while Maifushi, Kingdee International, and other related stocks declined by 5%.

On-chain data shows that the LAPTOP multi-sig address received 100 million tokens (10% of the total supply) roughly 7 days before the token’s launch. During the launch period, the address transferred and distributed approximately 42.5 million tokens: around 15.5 million to market maker GSR, and another ~14.5 million to an unlabeled address. The community has noted that this token allocation deviates significantly from the project’s disclosed tokenomics, sparking widespread questions. However, Hunter Biden responded this morning, stating he did not profit from the token issuance.

Dongcha Beating AI News Flash: DeepSeek is consolidating its models. The chat interface has merged its three original modes — "Fast", "Expert", and "Image Recognition" — into a single V4.1 Flash model that handles daily conversations, complex reasoning, and image understanding. The API segment is undergoing the same consolidation. The original V4 Flash and V4 Flash Vision Exp models have been taken offline, with their old model IDs temporarily reassigned to V4.1 Flash. Currently, only two main models remain: V4.1 Flash and V4 Pro. After 12:00 on September 14, V4 Pro will also be taken offline, with all requests routed to V4.1 Flash and billed at the Flash model’s rate. Until the V4.1 Pro model launches, DeepSeek will effectively rely on V4.1 Flash as its primary model for both chat and API services.