模型发布/更新
-
NVIDIA launches Nemotron 3.5 Lightning to accelerate local agent tasks
NVIDIA released Nemotron 3.5 Lightning, a customizable open-source 30B mixture-of-experts (MoE) model designed for…
-
SGLang announces Day-0 support for NVIDIA Nemotron 3.5 Lightning
SGLang announced Day-0 support for NVIDIA Nemotron 3.5 Lightning, an open-source model with a 30B total-parameter, 3B…
-
Ling-3.0-tiny goes open-source: how 1.3B activated parameters tackle real tasks
Ant Group's Ling open-sources Ling-3.0-tiny, a native hybrid reasoning model with 7.9B total parameters that activates…
产品发布/更新
-
ZCode upgrade adds Goal, Subagents, Remote Control and idle-time tasks
ZCode, deeply optimized for GLM, today launched four major features: Goal, Subagents, Remote Control and idle-time…
-
ChatGPT desktop app now supports importing work data from other agents
You can now keep work from other agents in sync with ChatGPT Work and Codex. Import projects, chats, skills and…
-
Databricks 开源 Metals v2:面向数百万行代码库的 Java 和 Scala 语言服务器
Databricks 开源 Metals v2,这是其面向数百万行代码库的 Java 和 Scala 语言服务器。Metals v2 专为智能体驱动的开发场景设计,旨在提升大规模代码库中的编辑与导航性能。目前 Databricks 的大部…
行业动态
-
Researchers find API flaw that can read encrypted reasoning traces of ChatGPT and other models
Alexander Panfilov's team found vulnerabilities in the APIs of major AI providers including OpenAI, Anthropic and…
-
Anthropic reportedly may go public as early as September this year, downplaying AI model competition to investors
Anthropic is in contact with potential investors to prepare for what could be the largest IPO in history, planning to…
-
Gemini surpasses 1 billion monthly users, becoming Google's fastest-growing product
More than 1 billion people now use @Geminiapp every month to spark new ideas and get work done. This is our…
-
NVIDIA reportedly developing trillion-parameter open-source AI model Nemotron 4 to challenge top global models
NVIDIA is developing a new generation of open-source AI models, Nemotron 4, with the largest expected to have at least…
-
NVIDIA 为何需要新供电架构以扩展 AI 算力性能
NVIDIA 主张以 800 VDC 直流配电替代传统交流多次转换,以降低损耗、支撑 AI 算力扩展。NVIDIA 与 Google、Microsoft 通过 OCP 联合制定该架构,已发布白皮书及 LVDC 固态变压器规范 v0.3,超…
-
英伟达循环融资达到新高度,黄仁勋是否过度出牌?
英伟达股价在相关消息公布后小幅下滑,收盘报217.55美元,盘前略有回升。金融记者Holger Zschaepitz和曾预测安然倒闭的Jim Chanos均对英伟达的循环融资做法表示担忧。此外,英伟达将发布真正开源(非仅开放权重)的Nem…
-
Electric 加入 Databricks,将 WASM Postgres 引入 AI 智能体沙箱
Databricks 宣布 Electric 团队加入,将 WASM Postgres 引入 AI 智能体沙箱。该技术让智能体在隔离环境中运行本地数据库,支持实时数据同步与离线操作,提升构建可靠、可验证智能体应用的效率。Electric…
论文研究
-
Unified Radix Cache: a single tree structure for hybrid model prefix caching
The LMSYS team proposes Unified Radix Cache, which uses a single token-keyed radix topology to uniformly manage FULL,…
-
AMIE research medical AI system demonstrates real-time clinical video consultations for the first time
Google Research and Google DeepMind advanced the medical AI system AMIE to conduct real-time clinical video…
技巧与观点
-
OpenAI's Astra model solves 10 math problems, exciting and worrying mathematicians
OpenAI announced that its unreleased Astra model solved 10 long-standing open math problems, spanning areas including…
-
Building a MiniMax-H3 multimodal video and audio generation pipeline with the ComfyUI API
This tutorial demonstrates how to use ComfyUI as a headless inference backend to build an end-to-end MiniMax-H3 video…
-
WeChat Xiaowei AI writing and AI comment features enter beta testing: the last human touch in Moments is fading
WeChat launched beta features for Moments AI writing and AI commenting based on Xiaowei. The former generates three…
-
Ryan Greenblatt: human-level AI could spawn runaway superintelligence via recursive self-improvement by 2032
Dwarkesh Patel and Ryan Greenblatt, chief scientist at Redwood Research, discuss the possibility of recursive…
-
A winner in every category: the valuation premium of SaaS leaders in the AI era
SaaS valuations are under pressure overall, but every sub-sector has produced an AI leader: CrowdStrike leads security…