AI 新闻 —— June 25, 2026
PulseAugur 当天浮现的 20 条头条故事 —— 综合实验室、论文及开发者社区的信号进行排序。
-
OpenAI to debut ChatGPT 5.6 with government-approved users first
OpenAI plans to initially release its new ChatGPT 5.6 model exclusively to customers approved by the U.S. government, according to a report by The Information. This staggered release, detailed in a memo from CEO Sam Altman, will be followed by a broader public release a few week…
-
DeepReinforce releases Ornith-1.0 open-source coding models that learn RL scaffolds
DeepReinforce has launched Ornith-1.0, a family of open-source coding models available under the MIT license. These models, built upon Gemma 4 and Qwen 3.5, are designed for agentic coding tasks and uniquely learn their own reinforcement learning scaffolds during training. The l…
-
Zhipu AI's GLM-5.2 challenges top closed-source models with open release · 1 source tracked
Zhipu AI has released its flagship open-source model, GLM-5.2, which supports a 1 million token context window and has demonstrated top performance in coding and long-range tasks. This release follows Anthropic's temporary withdrawal of its Claude Fable 5 and Mythos 5 models, po…
-
OpenAI unveils custom AI chip Jalapeño, Anthropic shifts agent paradigm · 1 source tracked
OpenAI has developed its first custom AI chip, codenamed Jalapeño, in collaboration with Broadcom. This chip is designed for LLM inference, powering services like ChatGPT and future agent products, with a focus on improving performance-per-watt and reducing reliance on external …
-
Anthropic accuses Alibaba of massive AI theft targeting Claude · 3 sources tracked
Anthropic has accused Alibaba Group of orchestrating a large-scale campaign to illicitly extract capabilities from its AI model, Claude. The alleged attacks, which involved over 28.8 million exchanges using nearly 25,000 fraudulent accounts, aimed to clone Claude's advanced reas…
-
ByteDance's Volcano Engine mass-produces Opus 4.6-level AI models
ByteDance's Volcano Engine has developed the capacity to mass-produce AI models comparable to Anthropic's Opus 4.6, but at a significantly lower cost. This advancement positions ByteDance to leverage AI across its diverse range of products and services. The company is reportedly…
-
OpenAI updates popular GPT-5.5 Instant model for refined responses
OpenAI has updated its widely used "GPT-5.5 Instant" model, aiming to refine its responses. This update focuses on improving the quality and precision of the AI's output. The specific enhancements are intended to make the model more effective for its users.
-
Claude Code adds safety guardrails for destructive Git commands
Claude Code has introduced a new safety feature that automatically blocks destructive Git commands, such as `git reset --hard` and `git checkout -- .`, when they are not explicitly requested by the user. This built-in guardrail operates as a classifier, inferring user intent rat…
-
Claude Code undergoes major LLM transformation; SK Hynix eyes NAND expansion
Claude Code has received a significant upgrade, with one expert noting it represents the third major transformation in Large Language Models (LLMs). This advancement is part of a broader trend in technology investment, where public fund management institutions are actively parti…
-
Anthropic accuses Alibaba of illicitly distilling Claude AI model
Anthropic has accused Chinese tech giant Alibaba of illicitly training its AI models using Anthropic's Claude. The AI lab claims Alibaba employed 25,000 fake accounts and conducted 28.8 million exchanges to "distill" Claude's capabilities, a process that significantly reduces de…
-
ByteDance's Doubao AI launches paid tiers amid high compute costs · 2 sources tracked
ByteDance's AI chatbot, Doubao, has introduced a paid subscription service with three tiers: 68, 200, and 500 yuan per month. This move aims to offset the substantial daily AI compute costs, which are estimated to be between 132 million and 240 million yuan. Doubao is a popular …
-
New App Enables Local, Offline Chat With Documents
Off Grid AI Desktop is a new, free, open-source application designed to enable users to chat with their documents locally on their personal computers. The tool handles the entire process, including embedding, vector storage, and language model inference, without sending any data…
-
Silidy Semiconductor secures funding for solid-state transformer modules
Silidy Semiconductor, a power semiconductor module manufacturer, has secured tens of millions of RMB in angel and Series A+ funding from Su Chuangtou and Jingjia Investment. The company, founded in 2023, specializes in full silicon carbide (SiC) power modules and IGBT+SiC hybrid…
-
Claude Code auto mode proves reliable for production with tests
A field report on Claude Code's auto mode indicates it is reliable for production code when used with proper safeguards, particularly a robust test suite. The author found that auto mode excels at tedious, mechanical tasks like refactoring and migrations, often completing them f…
-
Run Alibaba's Qwen LLM locally and offline with Off Grid AI Desktop
Off Grid AI Desktop is a new, free, open-source application that allows users to run Alibaba Group's Qwen large language models locally on their personal computers. This enables offline, private AI interactions, with the app supporting various Qwen model sizes and leveraging har…
-
Run Google's Gemma LLM Locally with New Open-Source App
A new open-source application called Off Grid AI Desktop allows users to run Google's Gemma language models locally on their Mac or Windows computers. This approach prioritizes user privacy by keeping all prompts and data on the user's machine, eliminating the need for cloud-bas…
-
Off Grid AI Desktop offers a GUI for local LLM use, rivaling Ollama
A new open-source application called Off Grid AI Desktop aims to provide a more user-friendly interface for running large language models locally on personal computers. Unlike Ollama, which requires command-line interaction and an API, Off Grid AI Desktop offers a graphical inte…
-
Developer boosts C LLM inference speed by 25x, hitting DRAM limits
A developer details the process of optimizing a C-based LLM inference engine, Project Zero, to achieve significantly faster performance on CPUs. Initially running BitNet b1.58 at 1.4 tokens/second, the project evolved over nine months to reach 36.25 tokens/second on a Xeon proce…
-
AI Coding Hackathon in Berlin to Redefine Software Engineering
A hackathon focused on AI coding tools is scheduled for July 15, 2026, in Berlin. The event aims to redefine software engineering by exploring the evolution of AI from simple autocomplete functions to collaborative coding partners. Participants will delve into the capabilities a…
-
New open-source app Off Grid AI Desktop offers local LLM chat and more
Off Grid AI Desktop is a new, open-source application designed to run large language models locally on a user's Mac or PC. It offers a similar download-and-chat experience to LM Studio but expands functionality to include on-device image generation, voice processing, and Retriev…