Binance adds 5 new TradFi perpetual contracts, offering up to 20x leverage.

On Friday, around $96 million in short-term call options flooded Sandisk (SNDK), Micron Technology (MU), Intel (INTC), and Marvell Technology (MRVL), with Sandisk surging over 11% at one point. All these options expire on October 2. Trading data shows roughly 10,000 call contracts for Micron, 4,200 for Sandisk, 20,000 for Intel, and 3,500 for Marvell, with total premiums amounting to around $96 million. CNBC host Jim Cramer posted on X that the trading pattern "looks exactly like Leopold is back," referring to AI investor Leopold Aschenbrenner and his fund Situational Awareness. According to regulatory filings, Situational Awareness’s two largest previous holdings were Sandisk and Micron, with positions of roughly $5.7 billion and $5.6 billion respectively at the end of June. On September 11, the Financial Times reported that Aschenbrenner had rebuilt positions in assets including AMD, Intel, SK Hynix, Sandisk, and CoreWeave via flex options. Nomura strategist Charlie McElligott noted at the time that roughly $315 million in option premiums had flowed into AI and semiconductor assets over several days. Aschenbrenner previously drew attention for his highly leveraged bets on AI and semiconductor stocks. His fund plunged 67% during the AI sector correction in July this year, after which he drastically cut positions. At the end of July, he wrote in a letter to investors that he would "learn the necessary lessons" and pledged that future public market investments would be managed on a "fully paid basis." To date, SEC filings have not disclosed the identity of the buyer of the nearly $100 million in options, so it remains unconfirmed whether the trades originated from Aschenbrenner. Unlike his previous use of total return swaps (TRS) for leverage, the market speculates he used fully paid options this time, with maximum theoretical losses limited to the premiums paid. However, the short term of the options means high time value decay and gamma risk.

Robinhood CEO Vlad Tenev stated that the company aims to evolve beyond a mere trading app, attracting users through trading products including stocks, cryptocurrencies, and prediction markets, then gradually converting them into long-term service clients for retirement accounts, investment advisory, banking, and wealth management. Tenev added that active traders serve as the "engine room and foundation" of Robinhood’s business: trading products help the firm acquire customers, while offerings like retirement accounts are more critical for long-term growth. As of the end of this year’s second quarter, Robinhood’s funded accounts rose 7% year-over-year to 28 million, with its average revenue per user (ARPU) jumping 24% year-over-year. Tenev also noted that over the coming decades, more than $100 trillion in global intergenerational wealth will transfer, and Robinhood seeks to expand its service suite to include trusts, estate planning, private wealth management, and private banking as its users’ wealth grows.

Insight Beating AI Express: Cua has open-sourced CUA-S1-FORMS, a model with just 706,000 parameters and 2.8MB in size, designed exclusively for automatic form filling. The system first extracts data like names and phone numbers from PDFs, then uses the model to determine which form field each piece of information belongs to, before handing off tasks to Cua Driver for execution. It follows the System One approach recently popularized by Jev, focusing solely on selection and judgment rather than text generation. While Jev targets more general decision-making tasks, CUA-S1-FORMS narrows the scope even further, addressing only information matching within forms. In Cua’s internal tests, CUA-S1-FORMS scored 99.7% on form-filling tasks, compared to Jev’s 83.6%. This result has a clear home-field advantage: the former was trained specifically for form filling, so it cannot be interpreted as a 700k-parameter model outperforming Jev overall. Jev aims to replace some large model decision-making in agents, while Cua is breaking tasks down even further. When tasks are sufficiently fixed, a general decision model is not even necessary.

According to Farside's monitoring, U.S. spot Bitcoin ETFs recorded a net inflow of $433 million yesterday, marking the second consecutive day of net inflows.

According to Arkham's monitoring, Zcash whale address t1Lyq transferred ZEC worth approximately $362.56 million last night, with $15 million deposited into Coinbase. This marks the first time the address has funded a cryptocurrency exchange in 10 months. Ten months ago, this batch of ZEC was valued at around $163.88 million. Arkham noted that the whale has generated an unrealized profit of roughly $361 million from holding ZEC since 2025.

Beating AI Express News: Jina AI has released its document parsing model jina-ocr-v1, which can directly convert PDFs, scanned documents, tables, and charts into Markdown. Post-trained on DeepSeek-OCR, the model retains the latter’s Mixture of Experts (MoE) architecture with approximately 3.4 billion total parameters, activating around 570 million parameters per token during decoding, and integrates FastMTP speculative decoding. In Jina’s internal tests, jina-ocr-v1 scored 91.14 on OmniDocBench v1.6, 0.89 points higher than DeepSeek-OCR-2; it also achieved 83.4 on olmOCR-Bench, a 7.4-point improvement over the DeepSeek-OCR base model it is built on. Throughput is one of Jina’s key selling points: on a single A100 with 32 concurrent requests, the model delivers 2.57 pages per second — the highest among the 14 systems Jina benchmarked, roughly 22% faster than DeepSeek-OCR’s 2.10 pages per second. The model weights are now available on Hugging Face under the CC BY-NC 4.0 license; commercial use requires contacting Jina.