Huang Renxun has left Beijing, and NVIDIA has not disclosed whether Huang Renxun met with the DeepSeek founder.

Beating AI News reports that Japanese developer Kamimoto has integrated Jev into MiniMax H3’s video generation workflow. Based on the operational status of each layer, Jev selects the proportion of Attention to retain from four preset options (1%, 3%, 5%, 10%), while Sparse Attention determines which specific computations can be omitted. When generating a roughly 5-second video on an RTX 4070 (12GB), the original 4-step approach took 6 minutes and 7 seconds; adding Sparse Attention plus Jev’s dynamic control cut the total runtime to 3 minutes and 34 seconds, a 41.7% reduction. However, this 41.7% speedup is not solely attributable to Jev. Sparse Attention itself can boost speed by skipping some Attention computations, and fixed retention rates of 5% or 10% also reduce runtime. Jev’s function is dynamic adjustment: allocating more computations to important layers and less to less critical ones. To date, there is no conclusive evidence that this dynamic adjustment delivers better video quality than fixed ratios, and the developer has explicitly stated that the quality has not been fully validated. The experimental code has been open-sourced.

According to market data from BIT (bit.com), NVIDIA’s stock valuation has been declining, raising warning signs about the chipmaker’s ability to sustain its explosive profit growth. NVIDIA’s current price-to-earnings (PE) ratio, based on estimated earnings over the next 12 months, is below 17 times, marking its cheapest valuation level in over a decade. This is only half of its 2025 forward PE (when NVIDIA’s revenue and profit growth was slower) and significantly lower than the over 25x forward PE it hit in May this year.

SEC officials stated that after the agency issued a conditional exemption order last week, the first batch of companies could submit applications as early as next quarter to operate tokenized U.S. stock trading platforms.

According to BIT (bit.com) market data, U.S. pre-market trading on Tuesday saw broad declines in the semiconductor, storage, and optical communications sectors. Semiconductor stocks: Lam Research (LRCX) fell 0.36%, Applied Materials (AMAT) dropped 0.41%, Arm (ARM) declined 0.1%. Storage stocks: SK Hynix (SKHY) fell 0.32%, Micron Technology (MU) dropped 0.4%, Western Digital (WDC) declined 0.93%, SanDisk (SNDK) rose 0.45%, Seagate Technology (STX) fell 1.06%. Optical communications-related stocks: Applied Optoelectronics (AAOI) gained 0.28%, Coherent (COHR) dropped 0.47%, Credo (CRDO) declined 0.79%, Lumentum (LITE) fell 0.63%, Nokia (NOK) dropped 1.1%.

Beating AI Express News: ByteDance has launched a Draft (rough cut) mode for the Seedance 2.5 API. Users can first generate a low-cost rough cut in 480P to confirm composition, actions, dialogue, and camera work; once satisfied, they can convert the selected rough cut into a 1080P final video. Jimeng has also integrated Seedance 2.5 (Draft Mode). This differs from the common approach of testing prompts in low resolution before re-generating high resolution with the same prompt. When generating 1080P, the model links to the original rough cut ID, reuses parameters including prompts, reference assets, seed, aspect ratio, and duration for re-inference, aiming to retain the original structure, composition, and actions as much as possible. Local textures, small text, and crowd count may still change, so it is neither a frame-by-frame replication nor a simple 480P upscaling to 1080P. Official calculations show that for a 5-second 1080P video (excluding video input), if users need an average of 4 generations to get a satisfactory result, Draft mode cuts costs by 57% and boosts overall generation speed by nearly double. If 20 generations are required to select a desired clip, cost savings can reach up to 77%. A similar Draft mode was previously available in Seedance 1.5 Pro. This update integrates the low-cost trial-and-error workflow into Seedance 2.5, enabling selected rough cuts to be converted into corresponding 1080P final videos. The most costly part of AI video generation is often the repeated trial runs, and this feature specifically targets reducing those expenses.

According to a Reuters report, a senior Iranian official stated that Tehran welcomes the resumption of diplomatic engagement if the United States takes concrete actions. The Iranian delegation has been fully authorized to advance the resumption of diplomatic contacts with the U.S. in New York. The delegation can discuss the specific details of an agreement to end hostilities with the U.S. via intermediaries in New York. If the U.S. eases military pressure and lifts port blockades, Iran can reopen the Strait of Hormuz within seven days.