Jensen Huang reportedly resigns as Nvidia CEO; CFO said to serve as interim CEO
Personnel ChangeAI Chips
Jensen Huang has reportedly resigned as CEO of Nvidia, with CFO Colette Kress said to be taking over on an interim basis until a successor is selected. Nvidia had previously been reported to be using an outside firm to search for a CEO successor and considering the appointment of an interim leader.
Google launches Gemini 4 Hydrogen, with tasks completed 2.5x faster
Model ReleaseDeveloper Tools
Google has launched Gemini 4 Hydrogen for everyday coding and agentic workflows. It completes tasks 2.5 times faster than Gemini 3.8 Flash. The model offers four levels of reasoning intensity, with API pricing of $1.50 per million input tokens and $7.50 per million output tokens. It is available through AI Studio, the Gemini API, and the Gemini app, among other channels. Antigravity's Orchestrator can combine Hydrogen and Carbon; in tests, this reduced quota consumption by 24% and increased task completion speed by 19%.
Inkling 2 Small weights released under Apache 2.0; vLLM adds inference support
Open WeightsModel Inference
Thinking Machines Lab has released the weights for Inkling 2 Small under the Apache 2.0 license. vLLM has also added inference support for the model. It uses a mixture-of-experts architecture with more than 500 billion parameters, accepts text and images, and supports a context window of up to 1 million tokens. In single-node tests on GB200 or GB300, it achieved output speeds of up to 630 tokens per second for a single user with four-way tensor parallelism.
Chip company launched by UBTECH and MetaX raises 500 million yuan in angel funding
FundingAI ChipsRobotics
Xixuan Chuangzhi, an edge-chip company jointly launched by UBTECH and MetaX, has completed a 500 million yuan angel round at a post-money valuation of 4 billion yuan. The company focuses on edge chips and computing platforms for embodied AI. Its first self-developed chip is planned as an SoC integrating a robot's brain and cerebellum, with an engineering validation platform for industry partners slated for 2027.
Edge0 opens weights for two 100B+ models aimed at local coding
Open WeightsOn-Device AI
Edge0 AI has released the open-weight models Edge0-145B and Edge0-120B for on-device coding tasks. According to the announcement, the models can run on laptops with about 29 GB of memory, decode at over 75 tokens per second, and support context windows of up to 256K tokens. They scored 86.5% on Terminal-Bench 2.1. The model weights are available on Hugging Face.
IBM introduces LLMON and two preview models to guard against prompt injection
Model SecurityModel Release
IBM has introduced LLMON, a markup language for language models that uses special-token delimiters to distinguish instructions from data, and released experimental 3B and 8B preview models based on Granite 4.1. Post-trained specifically for LLMON, the models achieved near-zero attack success rates on the OPI prompt-injection benchmark while maintaining performance comparable to Granite 4.1 Instruct.
Tether releases 800M-parameter BotPsy, optimized for on-device tool calling
On-Device AIModel Release
Tether AI Research has released QVAC BotPsy-0.8B, a model post-trained from Qwen3.5-0.8B for tool use and agent tasks on personal devices. It scored 44.0 on BFCL v4 and 58.7 on ACEBench, outperforming Qwen3.5-2B-Think on five of six agent benchmarks. The release includes GGUF files suited to local inference. In red-team tests, the unsafe-response rate was 21%.
Baidu launches Token Exchange, with more than 50 million daily calls
Model ServicesIndustry Data
Baidu has launched Token Exchange, a marketplace for large models. It verifies models through model-fingerprint matching, capability evaluations, and performance and availability tests, and assigns credit ratings. After one month of operation, the platform had onboarded more than 100 resource providers, with daily call volume exceeding 50 million and daily token usage exceeding 700 billion. Baidu also released the ERNIE Flash Think deep-reasoning model and the ERNIE Yijing micro-expression model, offering API services for both to enterprises.
Doubao Work has opened its platform to third-party models, adding Kimi 3.1 Pro, GLM 5.3, and DeepSeek V4.1 Flash. Users can choose a model for each task within the same platform. The model-selection page displays each model's quota-consumption multiplier. The platform has also introduced a limited-time promotion that reduces quota consumption for Doubao 2.1 Lite and DeepSeek V4.1 Flash.
World Labs' world model Atlas is now available to all users. Atlas uses a multimodal autoregressive diffusion Transformer architecture and supports dense reconstruction, sparse reconstruction, and text-only generation. It uses view prediction to unify generation and reconstruction tasks. The model ranked first on the Rapidata camera-motion benchmark.