NVIDIA has released Video Codec SDK 13.1, adding a set of video encoding and decoding upgrades for developers. The update includes AV1 hierarchical reference mode, a zero-copy transcoding architecture, frame-level precise search, enhanced hardware decoding statistics, and a new Docker development environment. According to NVIDIA, the release is aimed at improving performance and efficiency in AI video processing, streaming, and large-scale content delivery workflows. On the decoding side, SDK 13.1 adds per-macroblock decoding statistics for H.264 and H.265, while avoiding extra CPU parsing overhead. NVIDIA said demand continues to rise for high-quality video streaming, generative AI video tools, remote collaboration, and content distribution at scale. The company said the new SDK is designed to help developers build video processing pipelines that are more efficient, lower in latency, and easier to scale.

ByteDance said on July 31 that it has launched Seedance 2.5, a video generation model that can create a single high-quality clip up to 30 seconds long. The company said the model marks a breakthrough in handling longer-form narrative generation. Seedance 2.5 will be rolled out to Jimeng AI and the professional version of Doubao in phases. ByteDance also said API access will be connected to Volcano Engine Ark in the near term. The announcement was reported by PANews.

MiniMax on July 31 introduced H3, also branded as Hailuo 3.0, a new omnimodal generation model that the company said will have open weights released in the coming days. The launch follows MiniMax’s open-source release of its M3 text model in June and positions H3 as a system built to handle mixed text, image, video, and audio inputs in a single workflow rather than splitting those functions across separate models. According to the company, H3 can take text, images, audio, and video as inputs, including mixed prompts, and produce native 2K, 24 fps video clips running from 5 to 15 seconds with synchronized sound generated in the same pass. MiniMax said the price for 2K video generation is RMB 0.8 per second, or about RMB 12 for a 15-second clip, which it described as roughly one-third the cost of comparable flagship models. MiniMax also highlighted its Omni-Reference feature, which supports up to nine reference images, three reference videos, and three reference audio clips in a single generation. The model also supports text-to-video, image-to-video, multi-shot storytelling, instruction-based editing, and V2V Motion Transfer. Still, its broader standing remains unsettled: MiniMax’s claimed top ranking applies to video editing rather than overall video generation, while open weights have not yet appeared on Hugging Face.

A dense stretch of crypto news from July 26 to July 27 put regulation, platform risk, token supply and exchange transparency in focus. In South Korea, at least 29 offshore crypto derivatives exchange apps became unavailable for new installation through Google Play, expanding restrictions already affecting names such as OKX and Bybit. The report said users in Korea now see messages that the apps are unusable or unavailable in their region, while browser access and Apple App Store distribution remain unaffected. Since January, Google has applied region-specific support suspensions to virtual asset service provider apps that are not registered with the country’s Financial Intelligence Unit. In the U.S., the Commodity Futures Trading Commission’s Division of Market Oversight issued new guidance for prediction-market platforms including Kalshi, Coinbase, Polymarket and Crypto.com, warning against broad, template-based self-certifications for event contracts. The agency said each category must include detailed terms, settlement methods, data sources and compliance analysis. On the corporate and security side, WEMIX said a related contract’s ownership had been compromised, leading to the unauthorized minting of more than 5.22 million WEMIX, while BitMart’s publicly visible on-chain reserves fell sharply in the days before its shutdown announcement. Token Unlocks data also showed major scheduled unlocks next week for SUI, EIGEN, FF, KMNO and ENA.

Jacob Tsimerman, one of the latest Fields Medal winners, said at a press conference that he has already shifted toward AI safety research and will soon join OpenAI’s safety team, with OpenAI chief research officer Mark Chen publicly welcoming the move. The article frames that decision as part of a broader talent shift: leading AI labs are pulling in top researchers from mathematics, physics, biology, economics, and other disciplines, while firms such as Anthropic and OpenAI are building formal programs to embed non-AI experts directly into model development. It also highlights a parallel move in China. ByteDance’s Seed Edge team this week unveiled the Seed STEM Scientist Program, offering 100 collaboration seats for scientists and PhD students in mathematics, physics, chemistry, biology, and other frontier fields. According to the article, this is the first large-scale domestic initiative focused on using AI to accelerate frontier scientific discovery. The piece connects that launch with a broader argument: model competition is moving beyond leaderboard gains and benchmark scores toward longer-horizon bets on base architecture, evaluation systems, open-environment learning, and scientific research as a real-world testbed for the next phase of AI capability growth.

Users have spotted a beta version of Grok Imagine Agent Mode on the web, replacing the chat box with an infinite canvas that can generate full sets of images, videos, comics, and product assets from a single prompt.

Black Forest Labs unveiled FLUX 3 on Thursday, turning its flagship system from a still-image model into one that can also generate video with synchronized audio. The German AI lab said the model was trained jointly on images, video, and audio in a shared system, making FLUX 3 a multimodal model rather than a set of separate tools stitched together. Video is the main release feature. FLUX 3 can create clips up to 20 seconds long, with dialogue, sound effects, and ambient sound generated alongside the visuals and aligned to what is happening on screen. In early head-to-head preference tests, human reviewers picked FLUX 3 over Runway Gen-4.5 in 77% of comparisons and over Luma Ray 3.2 in 93%. Against Gemini Omni and Seedance, BFL said FLUX 3 held a narrower edge, winning 52% of evaluations. The company is also tying the model to robotics through FLUX-mimic, a system built with Zurich-based mimic robotics. BFL said Audi is already testing it for tasks such as fitting flexible door seals, and put end-to-end response time at about 101 milliseconds. FLUX 3 is not fully open yet: Video and Action are in early access through APIs and selected partners, while the open-weight Dev version for local use is scheduled for later in 2026.

HeyGen has launched Avatar V, an AI video tool that builds a photorealistic digital twin from a 15-second recording and supports lip-synced output in 175 languages across its paid plans.