US stock market opens: Nasdaq down 0.22%, Meta up around 3%

Beating AI Express News: After Ox Alpha was announced during the day, Zhipu AI officially open-sourced GLM-5.3-Flash in the evening. The model weights are now available on Hugging Face under the MIT license. With a total of 320 billion parameters, only 18 billion are activated per forward pass. It natively supports text, images, and video, making it the first native multimodal model in the GLM-5 series. According to the official, GLM-5.3-Flash outperforms GLM-5.2 overall, with its programming and agent evaluation scores approaching Claude Opus 4.8, while costing only one-tenth of GLM-5.2’s price. It also features a new base model, marking the first time a hybrid sparse and linear attention architecture has been introduced in the GLM main series, and is pre-trained on 30 trillion tokens of multimodal data. Earlier, the community’s small-sample DeepSWE test on Ox Alpha once yielded an 80% score, but this was based on only 10 questions. After expanding the test sample, the score dropped to around 63%, and the tester voluntarily corrected their initial claim. This score still places it in the first tier, though it is not as high as the initial 80% figure. With the model weights now open-sourced, developers can deploy it directly using frameworks like vLLM, SGLang, and KTransformers, eliminating the need to use APIs from anonymous models.

Aave Labs announced that its Aave mobile application will launch early access on August 26, with Ghost Passes being distributed to select users. These passes are invitation credentials, and their holders can gift them to other users, allowing recipients to bypass the waitlist and gain immediate access to the Aave App.

Pump.fun has announced that its application now supports HyperEVM. Users can trade any HyperEVM tokens directly with USDC, earn Callout rewards, and execute transactions at nearly zero fees.

On-chain analytics platform Bubblemaps published a report stating that Cosmos Labs previously disclosed a security vulnerability in its shared Cosmos EVM software, which has affected multiple blockchains using this module, including Nesa. The main attacker address 0x9AE7 first purchased approximately $250,000 worth of NES and bridged it to Nesa Chain, then exploited the balance vulnerability to inflate its holdings by 200 times, before bridging back around $50 million worth of NES to Ethereum. The wallet was initially funded with Monero-related funds; afterward, the attacker dispersed the tokens to multiple addresses, swapped NES for ETH via decentralized exchanges (DEXs), and deposited the proceeds into centralized exchanges. Due to the rapid withdrawal of liquidity from liquidity pools, most of the attacker’s trades experienced severe slippage. Bubblemaps noted that the attacker spent a total of roughly $255,000, ultimately receiving approximately $315,000 from the sale, resulting in a net profit of only about $60,000. NES’s price plummeted by 90% at one point, but has since recovered significantly, likely driven by arbitrage opportunities from price discrepancies between DEXs and centralized exchanges post-attack. Another Cosmos EVM project suffered an attack of around $1.4 million during the same period, but with distinct funding sources and operational methods, possibly carried out by a separate attacker.

Bloomberg ETF analyst Eric Balchunas has flagged another sign that anti-currency depreciation trades are replacing the AI frenzy in the market. Gold ETF GLD and BlackRock’s spot Bitcoin ETF IBIT have re-entered the top 10 ETF trading volume rankings, while semiconductor ETFs — which topped the charts all summer — have fallen in the standings. Balchunas noted that semiconductor ETFs’ trading activity remains above the normal level corresponding to their asset size, but has cooled significantly compared to earlier.

Beating AI News Flash: Recently, ChatGPT paid users have noticed that despite selecting the GPT-5.6 Sol Thinking or Pro models, their responses have suddenly become faster but lower in quality. Some users who captured network requests found that while the frontend requests GPT-5.6, the server actually returns responses from gpt-5-5-mini. Adam Fry, OpenAI’s ChatGPT product lead, has confirmed this is a routing bug. Approximately 3% of Pro and Thinking model requests were incorrectly routed to GPT-5.5-mini, and the issue has now been fixed.