Moonshot AI has released the weights of its native multimodal model Kimi K3.1 Pro. The model has approximately 2.8 trillion total parameters, activates about 104 billion parameters per token, and supports a 1-million-token context window. It scored 65.7 on Terminal-Bench 4.0, and its weights are available on Hugging Face. vLLM offers day-one deployment support across platforms including NVIDIA and AMD GPUs and Google Cloud TPUs.

Tencent Hunyuan has officially released Hy4, with approximately 770 billion total parameters and about 49 billion activated per inference. It supports a context window of approximately 1 million tokens. New capabilities include visual understanding and Computer Use, allowing it to operate software such as Blender and CAD through graphical interfaces and terminal commands. Hy4 scored 3.04 in an evaluation of 579 financial tasks. The model is available through Tencent Cloud, WorkBuddy, and other channels, with its weights planned for release later.

OpenAI says its autonomous AI workflow verified and patched 948 Linux kernel privilege-escalation vulnerabilities, at an API cost of approximately $1.2 million. Agents involved in the task also uncovered at least 10 Linux exploit techniques that experts had not previously known. OpenAI is working with Linux maintainers and Patch the Planet to advance responsible disclosure and remediation.

OpenAI announced plans to acquire Pierre Computer Company to expand Codex's code review and collaboration capabilities. Pierre's Code Storage and open-source tools such as Trees and Diffs support code storage, browsing, and review; some have already been integrated into Codex. The two companies plan to develop Git workflows that support parallel coding by agents, while existing open-source projects will remain open source. The transaction remains subject to regulatory approval and customary closing conditions.

Apple has opened Siri AI integration to developers in the iOS 27.3 beta. Developers can surface app content and actions through App Intents and use Skills to handle multi-step requests, combining natural-language capabilities with personal-context understanding. Submissions for eligible apps are expected to open on the App Store in spring; the feature is not currently available in the EU or to users under 13.

ChatGPT has reached 1.2 billion weekly active users, an increase of about 200 million in two months. Voice mode has 250 million weekly active users. ChatGPT's search volume is about 20% of Google's, and users generate approximately 1.5 billion images each week. OpenAI plans to combine chat and work modes.

Anthropic's biological research team worked with Claude to map the 3D structure of bacteria-derived Vault particles. The research found that related proteins in bacteria and their viruses can also assemble into these large, barrel-shaped structures. The team said Claude shortened structural analysis that would normally take weeks to months to just days, though researchers still had to correct errors along the way.

Hugging Face has launched RoboColosseum, a robot-model evaluation project offering access to more than 100 real robots across more than 140 real-world scenarios. The hardware includes Unitree G1, LeRobot SO-101, and Franka robots. The project offers 10,000 hours of free evaluation and will make run logs, including failure cases, available on Hugging Face Hub.

Pruna AI has open-sourced P-H3-Image, an image generation and editing model reworked from the video model MiniMax-H3. It reduces inference steps from 50 to 8 and model size by a factor of 7, supports output at up to 16 megapixels, and can edit up to 8 input images. The model has native support for ComfyUI and Diffusers.

NVIDIA Recovery automates fault handling for large-scale GPU clusters. It detects training-job crashes and hangs, as well as hardware and interconnect failures, then isolates nodes, restarts jobs from checkpoints, or stops them according to policy. Production data show an average detection time of 69 seconds and an average of 94 seconds to issue a recovery action. Full job recovery time still depends on scheduling and checkpoint loading.