Reflection AI 放出了首个开源模型 Beam,501B 只激活 23B,官方说打平 GLM-5.2 还省 3-4 倍算力,做编程和智能体的可以关注权重发布。
Latest
Nvidia 投资的 Reflection 出了首个开源权重模型 Beam,号称推理打平 GLM-5.2 还更省算力,权重这个月就放。
Reka AI 用 320 块 H100 三个月训出 19B 的 Rho-1,文本图像视频加机器人控制一个模型全包,算力开销比主流大模型小得多。
德国公司 Aleph Alpha 把 Kolibri 权重完全开源了,Apache 2.0 许可随便用,78B 参数 MoE 架构还特意喂了 21% 德语数据。
GPT-6 Sol Codex 近 30 万字系统提示词被扒出来了,反废话黑名单、60 秒汇报机制、心跳唤醒全在里面,做提示词工程的不看亏大了。
不生成文本、只做分类决策的 Jev 三周就跑到日处理 1 万亿 token,OpenAI 和 Cloudflare 都出了同路线产品,挺值得了解的新方向。
Cantina 把 Z.ai 的 GLM-5.3-Flash 微调成了专挖漏洞的开源模型,60 个私有漏洞任务解出 40 个,MIT 协议随便部署。
一篇把 Qwen 从 7B 到 2.4T 参数每个版本的发布和许可证变化都捋清的长文,还带了来源链接,想对比选型可以收藏。
Yandex 用一个 transformer 就把推荐系统的候选生成加排序全包了,Yandex Music 点赞数涨了 11.42%,做推荐的朋友可以看看。
四个前沿模型各有所长:Astra 擅长操作电脑,Argon 强在法律金融,Sol 写代码最便宜,看完就知道该用哪个。
Google 悄悄上线了 Gemini 4 Argon,主攻推理和安全方向,就是冲着 OpenAI 和 Anthropic 的旗舰模型去的,可以看看它到底强在哪。
NASA 和 IBM 把月球轨道器 17 年的观测数据训成了一个开源模型,预测极地冰误差能降 22%,做月球研究的可以直接拿来用。
AWS 开源了个 2B 小模型,专门替智能体做选择,115 毫秒出结果还带置信度,做路由和工具调用判断挺实用。
有人拿 GPT-6 Astra 把拿破仑一封 217 年没人读懂的密信破解了,6 小时搞定符号识别加密码分析,过程细节都在文里。
Cloudflare 用 Qwen 造了个决策模型,输出只按 schema 来,flash 版每百万 token 才 0.09 美元,比 Jev 便宜不少。
Google 把 Gemini 4 Argon 拿出来了,直接点名对标 GPT-6 Astra,想看两家旗舰怎么较量的可以看看。
This model enables local AI decision-making without cloud dependency, addressing privacy concerns for sensitive applications. Developers can implement it for real-time decision tasks in edge computing environments, with performance competitive to larger models while requiring minimal resources.
LEGO-Anything能将照片转成3D代码,但AI代理们对自己的重建结果毫无判断力,GPT-6 Astra准确率仅53%。
Black Forest Labs 推出 FLUX 3 Image,能精准排布 AI 元素,支持 4K 生成,比普通生图模型控制力更强。
Clef's integration of visual capabilities with extended context creates new possibilities for complex business decision systems. Developers can now build multimodal applications that process both text and images simultaneously, with Cloudflare offering customization options through their new reinforcement learning product. The model's performance on Jev benchmarks positions it as a strong alternative to existing decision-focused AI systems.
Gemini 4 Argon's focus on long-task execution and security defense positions it as a specialized tool for enterprise-grade applications rather than general use. The model's performance claims in code optimization and memory efficiency could significantly impact how organizations approach large-scale software development and migration projects.
This breakthrough shows how AI can transfer knowledge from human demonstrations to robotic systems without physical training, potentially accelerating robotics adoption in specialized fields. For researchers and developers in automation, this approach could reduce development time and costs for complex robotic tasks in manufacturing or laboratories.
微软新流式转录模型延迟仅 0.13 秒,错误率 2.5%,比 Grok 和 ElevenLabs 都低,实时体验更好。
Google 新出的 Gemini 4 Argon 模型,专精复杂领域,百万 token 输出,网络安全团队已抢先体验。
This open-source infrastructure democratizes access to MoE model training, potentially accelerating research beyond well-funded labs. Developers and researchers can now experiment with MoE architectures without massive resource requirements, though they'll still need significant GPU clusters for full-scale training.
Only the latest 2000 items are listed; for older ones see the daily archive