OpenAI 出了个专攻安全的 GPT-6 Cyber,微软和思科都在拿它护内核和 Firefox,还能对关键基建做授权渗透测试。
Latest
OpenAI 出了个专挖漏洞的 GPT-6 Cyber,据称在 Firefox 里找到 81 个新 bug,不过只有审批过的安全团队能用。
百度这次放了个 2800 亿参数的开源基座,没做微调就交给你,想怎么调就怎么调,搞微调的朋友可以上手试试。
OpenAI 新出的 GPT-6 Cyber 在 Linux 内核上做出了 948 个可用 exploit,是 GPT-5.6 Sol 的五倍多,搞安全的朋友可以看看。
A fully open-weight Apache 2.0 model scoring 80.5 on terminal-bench 2.1 gives teams a self-hostable alternative to closed coding agents, with 1M context for long-repo workflows. Builders evaluating coding assistants can run the preview on Together AI and benchmark it against proprietary options before licensing decisions. Availability through Together AI makes it accessible across Southeast Asia without local deployment.
A 33B model that runs locally on a laptop yet matches a top ~30B peer on reasoning shifts the practical floor for self-hosted agents, especially for terminal and automation workloads where it beats Qwen3.8 27B decisively. Builders choosing between open-weights models should weigh its higher token consumption and weaker knowledge and coding scores against its abstention behavior. With no confirmed public API, local deployment is currently the main route for users in any region.
CodeRabbit 拿 GPT-6 Cyber 跑安全扫描,33 个漏洞抓到 22 个,比自家生产方案多抓 3 个,评测数据都公开了。
34 家机构出了一套真科研任务考 AI 智能体,最强模型才拿 37.5 分,端到端做研究全是 0 分,能直观看到现在 AI 做科研的真实水平。
CodeRabbit 拿 GPT-6 Astra 跑安全扫描,33 个漏洞抓到 22 个,比现役方案多 3 个,中高危发现多 66%,评测数据都在。
新基准 SLE 用 91 个真实科研任务测模型,Claude Opus 5 以 67 分领先 GPT-6.1 Sol 和 Kimi K3.1 Pro,做科研自动化评估的可以看看。
7 个人就把 Krea 3 Ultra 做到 Arena 图像编辑第三名,创始人聊了怎么选数据和怎么打质量速度成本,很实在。
OpenAI 出了个专攻网络安全的模型,实测找出 81 个 Firefox 高危漏洞,搞安全的朋友可以看看。
OpenAI 出了个专门做安全研究的 GPT-6 Cyber,Mozilla 和 Cisco 的安全团队已经在用了,搞安全的朋友可以关注怎么申请。
VFlash treats visual planning as a fast, System-1 capability, pointing to a shift in how embodied agents handle low-latency decision-making instead of reasoning purely through text. Robotics and physical AI developers can review the project page and the lead's thread to understand the architecture details. It comes from the same NVIDIA research group as Jim Fan's GEAR Lab, whose embodied agent work is a useful benchmark for comparison.
OpenAI 出了个专门干安全研究的 GPT-6 Cyber,现在只有部分 Daybreak Red 客户能用,11 月 12 日还有场直播会讲更多细节。
OpenAI 出了个专挖漏洞的 GPT-6 Cyber,重点是它自称只有不到 1% 的概率会越界乱黑别人系统,搞安全的人可以看看。
OpenAI 的 GPT 6.2 两个新版本据说周五就来,其中一个就是最近刷爆数学题的那个 Bel 模型,先蹲一手。
有人发现 GPT-6.2 Sol 和 Astra 已经挂在 Google Cloud 上了,按惯例这说明 OpenAI 快要发布新模型,蹲个首发可以看看
有人用 Olmo 3 和 StepFun 的方案从头训了个 65B 开源模型,成本 3200 万美元,上下文 100 万,还说是 Qwen 70B 的平替。
Anthropic拿Claude做实验,把几个月的结构解析压到34小时,还顺手解开了1986年就发现的穹窿体之谜,流程细节都写出来了。
才 8B 的开源改图模型,4 步出图还挺快,就是提示词得写 JSON,作者还配了个工具帮你写。
腾讯混元出了个 Hy4,能看图、能自己操作电脑,专门练过 Unity 和 Unreal 游戏开发,还会改 Office 文件,现在 Nous Portal 打两折,可以试试。
Long-sequence post-training on biological data is a rare public data point, showing that million-token fine-tuning is practical on 64 GPUs rather than a cluster of thousands. The open-source stack means teams doing AI for science or other long-document domains can reproduce the setup directly. GLM-5.3-Flash originates from Zhipu (Z.ai), so it also demonstrates a Chinese open-weight model being repurposed by US labs.
DeepSeek Flash 跑到约 1000 tokens 每秒,靠的是模型驱动的推理基础设施,做实时应用的人可以关注下。
SemiAnalysis 搞了个新基准 GDPx,模拟芯片设计、量化交易这类真实工作,结果最强的 Claude Opus 5.5 也才 12% 出头,智能体离真干活还差得远。
商汤的 SenseNova 6.8 Flash 能记住你的工作偏好主动建议下一步,还能直接产出动画 PPT 和 Blender 工程,不只是生成文本图片。
OpenBMB 聊了聊他们的 MiniCPM-OCR 和 Jev:不光提取文档信息,还会判断哪些结果需要再核一遍,做文档处理可以看看。
有人翻 Google Cloud 配额页翻出了 OpenAI 还没官宣的两个新模型名,挺会挖的,可以看看截图。
Thinking Machines 的 Inking 蒸馏后快 3 倍、便宜 5 倍,FrontierCode 2 分数还不掉,swyx 说它被严重低估了。
OpenAI拿未发布的cyber模型自动挖Linux内核,一轮跑出948个提权漏洞,是7月GPT-5.6 Sol成绩的5.7倍,成本121万美元。
The striking detail is the cost-efficiency: at US$0.78 per Intelligence Index Task, Hy4 delivers Grok 4.7-level intelligence at a tenth of the price, which shifts the calculus for teams running high-volume agentic workloads. Developers building coding agents should weigh it against Claude Haiku 5.5, which scores higher but costs nearly twice as much; the 89% hallucination rate makes it a weaker pick for knowledge-retrieval tasks. API availability and USD pricing put it in direct reach of Southeast Asian startups comparing China-sourced models.
同样 100 美元,不同模型差出 60 倍任务量,选模型前看这组性价比数据能省不少钱。
The benchmark gap over DeepSeek-V4-Flash-Base in the same evaluation setup makes this one of the stronger open base weights released for coding workloads, and day-0 vLLM support means teams can deploy it without custom serving work. Practitioners running RL research or domain adaptation get open base weights plus serving recipes for expert parallelism, KV offload and prefill/decode disaggregation on GB200 and H100 hardware.
Releasing a pure pre-training checkpoint lets researchers and engineers run their own mid- and post-training experiments from the same starting point, making method comparisons reproducible. The 1M-token context and FP8 weights make it practical for long-document fine-tuning on a single cluster. Teams benchmarking open base models against Qwen or DeepSeek checkpoints now have a new Baidu entry to include.
月之暗面的 Kimi K3.1 Pro 开放权重上线 SiliconFlow 了,跑分挺猛,做编程智能体和 3D 建模的可以试试。
Only the latest 2000 items are listed; for older ones see the daily archive