MiniMax H3 Multimodal Video Model Expected to Release Weights Today
要闻
MiniMax is expected to release the model weights of its new-generation multimodal video model MiniMax-H3 today. As of this writing, the relevant page still shows it is about to be open-sourced.
MiniMax is about to open-source its next-generation open multimodal video model MiniMax-H3. The related page shows that the expected release time is August 3, 2026, 06:00 (UTC+08:00), when users can obtain the model weights on ModelScope. However, as of the time of this release, the page still shows it as pending open-source.
FFmpeg Thanks Cursor AI for Providing Free Credits to Multiple Developers
行业动态
FFmpeg's official account posted to thank Cursor for providing free credits to multiple FFmpeg developers. FFmpeg said these credits will be used to support the project's ongoing development and code review work.
The official FFmpeg account publicly posted to thank Cursor AI for providing free credits to multiple FFmpeg developers. According to FFmpeg, these credits will support the continuous work of the FFmpeg project, including development and code review. FFmpeg expressed its gratitude to Cursor for its support of FFmpeg and its community.
Former Google Team Member Says Google Had a Similar Product Called LMChat Before ChatGPT
行业动态
Tibo, head of Codex, recently responded to netizens saying that Google had a similar product called LMChat about a year before ChatGPT's release. He worked on the relevant team, but the product was not launched due to concerns about impacting the search business.
Codex lead Tibo recently responded to a netizen, saying that Google had developed a similar product about a year before ChatGPT's release. He had worked on the relevant team, initially codenamed LMChat, and later changed to another codename. Tibo said Google did not release it because it worried the product would impact its search business, and DeepMind was also prevented from launching products that could disrupt Google. Earlier, a netizen cited an interview with Jeff Dean claiming that Google had an internal chatbot before ChatGPT.
Andrej Karpathy Says We Should Stop Testing LLMs with Pelicans Riding Bicycles
技术与洞察
Andrej Karpathy recently posted that we should no longer test LLMs with pelicans riding bicycles. He conducted a 3D world generation capability test using Opus 5, inputting the first paragraph of "The Lord of the Rings" and asking it to render the story using three.js. He said the result was "a bit rough but interesting."
Andrej Karpathy recently posted that we should no longer use a pelican riding a bicycle to test LLMs. He used Opus 5 to conduct a 3D world generation capability test, input the first paragraph of The Lord of the Rings, with a budget of about 1 million tokens (about $10), and asked it to render the story using three.js. Opus 5 ran for about 2 hours, wrote 5,500 lines of code, and procedurally rendered the story scene. Karpathy said the result was rough but runnable; the LLM needs to place and arrange polygon assets in a 3D coordinate system and write animation code. Karpathy also pointed out that current LLMs have weaknesses in the world and game domains, cannot efficiently natively perceive video or play games in-game, and need to slowly take screenshots to check work.
Qwen et al. Jointly Release Native Computer Use Agent Qwen-CUA
模型发布
The Qwen team, together with XLang Lab, released a native Computer Use agent called Qwen-CUA. The agent perceives the interface only through screenshots and operates the computer using keyboard and mouse. The technical report and reference demo are now available.
Qwen Team and XLang Lab jointly released Qwen-CUA, a native Computer Use model and agent based on Qwen. Qwen-CUA perceives the interface state solely through screenshots, without relying on DOM trees, accessibility metadata, or task-specific APIs, but executes operations in browsers, desktop applications, and professional software through native keyboard and mouse events. The model is based on a 397B-A17B mixture-of-experts architecture, scoring 86.2 on the OSWorld-Verified benchmark, and the larger version Qwen-CUA-Max has more than one trillion parameters, reaching 87.6 on the same benchmark. Currently, only the technical report and reference demo are available on GitHub.
Note: Content is AI-assisted and may contain hallucinations and errors.
