Chinese contentGeneralist AI发布GEN-1.5：从3-12秒演示中学习新任务的机器人基础模型
Generalist AI发布了GEN-1.5，一个机器人基础模型，只需3-12秒的演示就能学习新任务，比以往模型更高效。
Latest
Every published item, in order · 2490 items
Aug 24
Thomson AI model by Reuters
Thomson offers a cost-effective alternative to US models for document review, with customizable features at lower long-term costs than OpenAI or Anthropic.
IT之家Source
Google Research ME-POIs Framework
A mobility-only variant of ME-POIs outperformed Gemini text embeddings on price-level classification, offering better location intelligence for businesses and urban planners.
Samsung's HBM Vision Beyond Memory
Samsung's aHBM could increase GPU compute density by 5-10%, potentially improving AI training efficiency in ASEAN data centers.
IT之家Source
WRC 2026 native multimodal world model
The model's multimodal capabilities could enhance AI applications in ASEAN's gaming and simulation industries.
Chinese content月之暗面Kimi K2.5月底结束服役
月之暗面Kimi K2.5模型月底结束服役，是之前最智能的模型，支持多模态架构，和Kimi K3相比，Kimi K2.5在通用智能任务上表现优异。
IT之家Source
Alibaba DAMO LiON liver cancer model
DAMO LiON helps detect small liver tumors missed by doctors, potentially saving lives in China and Southeast Asia where liver cancer prevalence is high.
IT之家Source
Tomorrow 08:00, the daily lands in your inboxTelegram channel or email. Same screening as this page, no ads.
Moore Threads releases MTT S5000 Prefill-as-a-Service whitepaper
The solution offers cost-effective inference for long-context applications like AI agents and document analysis, potentially lowering operational costs for businesses in the region.
IT之家Source
iFLYTEK Optimizes Architecture for Compute Constraints
iFLYTEK's Spark X2-Flash on Huawei Ascend 910B offers a viable alternative when Nvidia H200 is unavailable in China.
Chinese contentHarvey推出Harvey Tenet：基于Kimi K3的微调模型，用于长期法律代理工作
Harvey发布了Harvey Tenet，这是基于Kimi K3的微调模型，适用于长期法律代理工作，比其他模型效率更高。
Aug 23
Zhiyuan Elf G2 Wins Two Golds
Elf G2 demonstrates practical applications in industrial and service scenarios, competing favorably against Boston Dynamics' Atlas robot in real-world tasks.
IT之家Source
Models18:18
Chinese contentAI成自身最大客户，OpenRouter上代理令牌使用量激增14倍OpenRouter上AI代理令牌使用量激增，成为自身最大客户，了解代理令牌的消耗情况，对比人类使用量增长。
Chinese contentOpenAI 投资公司转用中国开源模型打造法律 AI
Harvey 用 Kimi K3 模型打造法律 AI，比 Fable 5 和 GPT-5.6 Sol 强，成本还低。
IT之家Source
Chinese contentAnthropic Claude 模型安全漏洞
Anthropic 的 Claude 模型存在安全漏洞，了解旧模型的风险和 Anthropic 的应对措施很重要。
IT之家Source
Chinese contentInherent发布Faraday智能体，复现能力超GPT-5.5与Claude Opus 4.8
Inherent新推Faraday智能体，小模型大能力，强化学习助其超越GPT-5.5与Claude Opus 4.8，值得关注。
IT之家Source
Chinese contentInherent AI团队超越Anthropic和OpenAI
Inherent的Faraday AI团队刚刚在复现研究论文方面超越了Anthropic和OpenAI，值得一试！
Aug 22
Chinese content神秘Ox Alpha AI模型限免上架：DeepSWE跑分碾压Claude Fable 5
Ox Alpha模型免费限免，DeepSWE跑分领先，多模态输入，与智谱架构相似，值得一试。
IT之家Source
Models09:33
Chinese contentMichael Polansky AI训练活体皮肤模型Lady Gaga的伴侣Michael Polansky开发的AI技术，能培养活体皮肤，用于发现新护肤成分，比传统方法更高效。
Chinese contentDeepseek发布实验性Flash视觉模型，在智能体基准测试中与Opus 4.8匹敌
Deepseek新推出的V4-Flash-Vision-Exp模型，将图像理解能力与文本处理结合，在多模态智能体基准测试中表现优异，与Opus 4.8不相上下，值得一试。
Amazon Bedrock Query-Aware Compression
This method reduces RAG costs on Amazon Bedrock without compromising answer quality, making it valuable for cost-conscious developers in the region.
Aug 21
DeepSeek V4-Flash-Vision-Exp model launched
The model offers competitive multimodal capabilities at existing pricing, with a free Files API option to reduce bandwidth costs for developers in the region.
IT之家Source
Chinese contentQwen3.8-27B新基准：本地Opus 4.6横扫消费级硬件
开源开发者有了新基准，Qwen3.8-27B横扫消费级硬件，比肩GPT-5.6 Luna和DeepSeek V4 Flash，参数量27亿，4位量化后适配24GB GPU，超越它才是新标准！
Chinese content商汤开源8B多模态大模型SenseNova U1.5 Lite
商汤新开源的SenseNova U1.5 Lite模型，支持4K图像输出，功能强大，是处理复杂视觉任务的利器，值得一试。
IT之家Source
Chinese content阿里云：GLM-5.3、Deepseek-V4-Pro等模型上线千问AI平台
阿里云新上线GLM-5.3和Deepseek-V4-Pro，编程和智能体应用场景更强大，值得开发者关注。
IT之家Source
Ant Group open sources Ling-3.0 model checkpoints
The open checkpoints allow researchers in ASEAN to continue training these models for local applications without starting from scratch.
IT之家Source
Liquid AI Releases LFM2.5-DSpark Models
The models offer significant speed improvements for developers using LFM2.5 in real-time applications.
Only the latest 2000 items are listed; for older ones see the daily archive