SponsoredAI 会员在线充值
Visit sponsor Chatgpt/Claude/Grok 会员24h在线自助充值 安全快速 稳定保障
奥创(Ultra)商城会员充值中心Loading…
SponsoredChatgpt/Claude/Grok 会员24h在线自助充值 安全快速 稳定保障
奥创(Ultra)商城会员充值中心
Sponsored本页公开展示实时样本、模型覆盖与响应时间。先用 5 元按量跑自己的 Claude Code/Codex 任务，成功后再决定是否长期使用。
ACS Gateway
Sponsored【腾讯云】WorkBuddy 全场景 AI 办公工作台，团队旗舰版低至198元/月
腾讯云
Sponsored人人可用的AI应用创作平台 自然语言交互，全栈应用搭建，伟大创意快速实现;个人版新用户注册送5000积分
阿里云Source-reviewed updates on AI models, APIs, research, security, and regulation. Read original evidence and track pricing and model changes.
Moonshot AI’s Kimi K3 is now available through Fireworks for hosted deployment and LoRA fine-tuning. The model has roughly 2.8 trillion parameters, a 1-million-token context window, and native vision support.
Baseten has made Moonshot’s Kimi K3 available through its Model APIs as a day-one launch partner.
Moonshot AI's Kimi K3 is now available on Modal through its Shared API and Auto Endpoint offerings. Modal says it added a custom DFlash speculative model tuned to K3 and can deliver up to 460 tokens per second.
Moonshot AI has released MoonEP, an open-source communication library for distributed Mixture-of-Experts training and inference. It targets communication overhead in expert parallelism with dynamic redundant experts, online planning, and zero-copy data movement.
Moonshot AI’s Kimi team and kvcache-ai have released AgentENV, a distributed platform for running agent environments at scale. Its components were used in agentic reinforcement-learning training for Kimi K3.
Moonshot AI has released FlashKDA, a CUTLASS-based implementation of Kimi Delta Attention kernels. The company reports 1.72×–2.22× faster prefill than the flash-linear-attention baseline on NVIDIA H20 GPUs.
Moonshot AI has released the weights and technical report for Kimi K3, a 2.8-trillion-parameter MoE model with native visual understanding and a 1-million-token context window.
DeepSeek has opened its official V4-Flash API to public beta, reporting stronger agent performance and support for the Responses API format and Codex.
DeepSeek says the current 75% discount for its V4-Pro API will remain in effect after the promotion ends on May 31, 2026.
DeepSeek has released the V4 Preview, featuring two Mixture-of-Experts models with open weights, million-token context windows, and immediate API and web access.
DeepSeek has released V3.2 as the successor to V3.2-Exp across its app, web product, and API. V3.2-Speciale is a higher-compute reasoning variant currently available through the API for evaluation and research.
DeepSeek says earlier versions of the DeepSeek-V3.2-Exp inference demo used inconsistent RoPE input layouts in the indexer module, which could reduce performance. The updated demo code resolves the issue.
DeepSeek has released the experimental V3.2-Exp model, built on V3.1-Terminus and introducing DeepSeek Sparse Attention for more efficient long-context training and inference. The model is available across the app, web, and API, with API prices reduced by more than 50%.
OpenAI is inviting feedback on Codex, including small changes that could make routine use more efficient and reliable.
OpenAI engineer Jason Liu says he created a video with Codex to describe the company’s mission and what it feels like to work there. He shared it internally before publishing it publicly with team support.
OpenAI is reducing GPT-5.6 Luna pricing by 80% and GPT-5.6 Terra pricing by 20%. The lower rates also affect usage accounting in Codex and ChatGPT Work, while GPT-5.6 Sol gets a faster API option.
OpenAI is launching ChatGPT for Academic Researchers, starting with 10,000 scientists, mathematicians, and engineers and scaling to 100,000 by 2027.
OpenAI reset usage limits for ChatGPT Work and Codex users after GPT-5.6 Sol consumed allowances faster than expected during demanding workflows.