模型发布/更新
-
OpenAI's largest ChatGPT voice upgrade: bidirectional AI voice model Bidi 1 now in testing
On June 23, some users reported that the bidirectional AI voice model Bidi 1 appeared in the ChatGPT web and app…
-
Qwen-AgentWorld 开源:让 Agent 学会“先预测,再行动”
通义千问推出首个原生语言世界模型 Qwen-AgentWorld,覆盖 MCP、Search、Terminal、SWE、Web、OS、Android 七大领域。模型基于超 1000 万条真实交互轨迹,经 CPT→SFT→RL 三阶段训练,…
-
New GPT-5.5 Instant version, conversations more fun
We bring you a new version of GPT-5.5 Instant, and it is now much more fun to chat with. Our most-used model now…
产品发布/更新
-
Doubao officially launches Professional Edition today
Doubao Professional Edition is now available, built on the Doubao 2.1 series of large models and aimed at complex…
-
Figma bets on human judgment at Config 2026, but canvas AI capabilities come from third parties
At Config 2026, Figma extended its design canvas to code, animation, 3D depth, and shader effects, and integrated the…
-
Volcano Engine launches Agent Ready infrastructure, upgrades AgentKit and ArkClaw Enterprise Edition
At the FORCE conference, Volcano Engine launched Agent Ready infrastructure for enterprise intelligent applications,…
-
Notion uses Cursor SDK to embed coding agents
Notion embedded coding agents into its product within weeks via the Cursor SDK. Users can @Cursor in documents,…
-
Perplexity launches Computer for Counsel
Launching Computer for Counsel. Computer now connects to the research databases, document tools, and case management…
-
Mistral AI 为 Connectors 推出多项安全与可控新能力
2026年6月24日,Mistral AI 发布 Connectors 多项新能力:Enriched admin controls(GA)支持按工作空间设置连接器访问权限并单独开关工具;API keys with connector sc…
-
FFASR leaderboard released: ASR evaluation under real far-field conditions
Treble Technologies and Hugging Face jointly launched the FFASR (Far-Field ASR) leaderboard, the first open-source…
-
OpenAI and Broadcom jointly release LLM inference chip Jalapeño
OpenAI and Broadcom released Jalapeño, their first self-developed inference accelerator, designed from scratch for…
论文研究
-
DFlash: block diffusion draft model delivers up to 15x throughput gain
DFlash, proposed by the UC San Diego team, is a lightweight block diffusion draft model for speculative decoding. It…
-
Thinking is recalling: how reasoning unlocks parametric knowledge in LLMs
Google Research found that reasoning (chain-of-thought) helps large language models (LLMs) recall simple facts, even…
技巧与观点
-
ByteDance VP Hong Dingkun: AI Coding practices and exploration
At the Volcano Engine Force conference, ByteDance VP Hong Dingkun shared AI Coding practices. Over the past year,…
-
AI expected to replace engineering jobs, but new data shows engineering was the most resilient role in 2025
SignalFire, a venture capital firm, tracked data on millions of employees at 80 million companies and found that…
-
OpenRouter zero data retention (ZDR) practice: 97 new models, traffic share near half
OpenRouter's zero data retention (ZDR) guarantee that user prompts and model responses are not stored, with metadata…
-
NVIDIA NeMo AutoModel: one line of code accelerates Transformer MoE fine-tuning
NVIDIA NeMo AutoModel is an open-source library based on Transformers v5, adding Expert Parallelism, DeepEP fused…
-
MiniCPM-V 4.6 runs at high speed on Apple Core AI
🥳 Thanks for sharing, @MLBoy_DaisukeMajima 🚀 MiniCPM-V 4.6 running on-device at this speed is truly impressive —…