Mistral launches Mistral Large 4 preview, a 1T-parameter open-weight model
Read original source →mistral.ai
Mistral Large 4 ("Le Chonk") is a 1 trillion-parameter natively multimodal model with 49B active parameters, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters. Preview API is live on Mistral Studio; weights drop by end of month after red-teaming with cybersecurity partners.
1 source
More stories today
Sequoia post recounts Cresta's path from Stanford lab to Salesforce
Cresta, founded in 2017 out of the Stanford Artificial Intelligence Laboratory by Zayd Enam with Tim Shi, Sebastian Thrun and Andre Esteva, is joining Salesforce. It emerged from stealth in 2020 with AI that coached contact-center agents in real time.
Sequoia Capital·1 hour ago

OpenAI details ChatGPT and Codex use in physical-world engineering
OpenAI shared progress on ChatGPT and Codex aiding mechanical design, electrical engineering, and simulation, including faster iteration for its own hardware teams. It is also testing these workflows with industry collaborators across CAD and PCB design.
OpenAI Blog·1 hour ago

Bender: Anthropic watermarking only now due to EU AI Act
Emily M. Bender·1 hour agoMeta seeks up to $25B in bond sale to fund AI buildout
Meta is marketing US dollar bonds for the first time in six months, aiming to raise $20B–$25B in the US and Europe. The debt sale funds infrastructure and AI models and services as tech giants brave investor skepticism over the AI buildout.
CNBC Technology·1 hour ago

OpenAI says ChatGPT search volume hits 20% of Google's
ChatGPT now processes one-fifth the search query volume of Google, per OpenAI consumer product lead Adam Fry, who announced the figure at a Monday press event. Fry did not disclose the source of the Google volume number; Bing held about 3% share after a decade, per Satya Nadella in the 2023 antitrust trial.
The Verge·1 hour ago

Bloomberg Opinion questions how AI giants will pay for compute
Parmy Olson's Bloomberg Opinion video examines how Anthropic and OpenAI will fund vast computer warehouses given few productivity gains so far, and critiques the justifications she's hearing from Silicon Valley executives.
Bloomberg Technology·2 hours ago

Replicate compares MiniMax H3 video models by price
A 10-second 768P clip costs $0.60 on H3 Highspeed, $0.40 on H3 Lite, and $0.80 on H3; at 2K an H3 clip runs $1.30. All three accept 4-15 second clips with text-, image-, video-, or audio-guided generation.
Replicate Blog·2 hours ago

Hugging Face releases CWE-bench v2 with more coverage and transparency
CWE-bench v2 expands benchmark coverage and quality over the original, with the team emphasizing transparency in how cases are selected and scored. The post frames it as a broader, higher-quality basis for evaluating AI on real-world CWE vulnerability classes.
Hugging Face Blog·2 hours ago