Ad
Short

xAI has released Grok 2 as an open model, including the weights. Elon Musk announced on X that Grok 2.5, xAI's top model for 2024, is now open source. The weights for Grok 2 are available on Hugging Face. Musk also said Grok 3 will be released as open source in about six months.

Grok 2 is available under the xAI Community License. Usage is free for research and non-commercial projects, while commercial use must follow xAI's guidelines. The license prohibits using Grok 2 to develop or train other large AI models. If you redistribute the model, you have to credit the source and include "Powered by xAI."

Ad
Ad
Short

Open-weight reasoning models often use far more tokens than closed models, making them less efficient per query, according to Nous Research. Models like DeepSeek and Qwen use 1.5 to 4 times more tokens than OpenAI and Grok-4—and up to 10 times more for simple knowledge tasks. Mistral's Magistral models stand out for especially high token use.  

Average tokens used per task by different AI models. | Image: Nous ResearchIn contrast, OpenAI's gpt-oss-120b, with very short reasoning paths, shows that open models can be efficient, especially for math problems. Token usage depends heavily on the type of task. Full details and charts are available at Nous Research.

High token use can offset low prices in open models. | Image: Nous Research
Short

This weekend, Google is giving users three free video generations with its AI video tool Veo 3 in the Gemini app. Veo can create short AI videos with sound and is currently the most realistic video model on the market. The promotion runs until Sunday, August 24, at 10:00 p.m. PT.

A humorous 8-second short video portraying a community theater-style play about AI video generation overheating Google's AI chips. | Video: Veo 3 prompted by THE DECODER

Normally, Veo is only available to paid Gemini users, starting at around $20 per month, or through the API for about 50 cents per second. Google could be using this promotion to test the system's stability ahead of a wider release. Since Veo launched, users have generated millions of videos, according to Google, though this activity isn't mentioned in the company's latest AI energy report.

Ad
Ad
Short

Nvidia has stopped producing H20 and 700,000 AI chips intended for China are now sitting idle.

After a temporary green light from the US government, Nvidia had promised Chinese customers about 700,000 H20 AI chips. These chips are stripped-down versions designed to meet US export rules, making them legal for the Chinese market. Now, a new directive from Beijing is forcing local companies to stop buying Nvidia chips over security concerns. As a result, thousands of finished chip dies are sitting unused at Amkor, a US-based packaging partner. The supply chain has ground to a halt, even though Washington and Nvidia had already reached a political agreement. The situation highlights how AI hardware is increasingly caught in the middle of geopolitical tensions. Earlier reports suggested the US is adding tracking chips to AI hardware bound for China.

Ad
Ad
Short

Dynamics Lab has launched Mirage 2, the latest version of its generative game world engine. With Mirage 2, users can upload their own images - like sketches or photos - and turn them into interactive game worlds. The engine also lets players change the game in real time by typing in commands. Worlds can be saved and shared. While Mirage 2 makes clear technical gains over its predecessor, it still struggles with precise controls and visual stability. In both areas, Genie 3 from Google DeepMind is far ahead, but Genie 3 isn't available yet and likely requires much more computing power. A Mirage 2 demo is available online.

Google News