Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding
Ripple-65B Preview 在 TerminalBench 2.1 上取得 80.5 分,成为美国开源权重模型在 Agentic Coding 上的新 SOTA。
Ripple-65B Preview 在 TerminalBench 2.1 上取得 80.5 分,成为美国开源权重模型在 Agentic Coding 上的新 SOTA。
Kimi K3.1 Pro is now open-weight: a native multimodal agent model for text and images with a 1-million-token context window. It uses a MoE architecture with 2.8T total parameters and 104B active, post-trained from Kimi K3 with large-scale joint reinforcement learning. Moonshot AI says coding, 3D creation, financial research and factual accuracy improved.
Why it matters: The release sets out what multi-turn tool calling must send back, so integrations need to keep full reasoning content and tool-call records to hold the conversation context.
OpenAI 称其硬件团队用 GPT-6 Astra 在不到一周内完成了一款可与 Codex 编码智能体对话的设备设计,作为业余项目推进。Astra 负责外壳、PCB 布局、元件选择、走线、固件和用户界面,代工厂在不到一个月内生产出小批量原型,但因模型未总能考虑制造约束、公差和装配细节,设计经历了多轮修订。
Simon Willison 在笔记本上运行 Qwen 3.8 27B,仅凭原始权重、无工具调用,成功识别三个游戏网络的用途。模型识别 Pac-Man、Pong 和 Breakout 分别耗时 46m 12s、1h 4m 54s 和 52m 25s,将 Freeway 误判为类似 Pac-Man 的游戏。
建筑师测试13个AI模型后认为,Claude Opus 5.5在专业指导下可为简单建筑生成达到建筑许可质量的图纸。模型通过Python代码生成可缩放、可在CAD中编辑的矢量线条,住宅图纸耗时十二小时、计算成本数百美元。各模型获得的人工反馈量不同,Opus 5.5接受了最多反馈,包括两轮红线批注,最终图纸仍遗漏当地法规要求的第二条疏散路径。
LongCat-2.5 是面向长程智能体任务的原生多模态模型,总参数量为 1.6 trillion,每个 token 约激活 48 billion 参数,在 LongCat-2.0 基础上强化视觉理解与智能体编码。
一个据称是 Anthropic 下一代 Mythos 级训练任务 numbat 的预训练中间检查点被泄露,规模约 14T 参数、114 层,采用混合 GDN/MLA 与 MoE(1,224 专家,top-48 + 1 共享)架构及 NVFP4 格式。
Memdoor 复现 GVS5H 多智能体流程,五个 Qwen3.8-27B 在前十道 LiveCodeBench Hard 题目上测试五轮,平均解出 9.8 题,原作者流程为 9.6 题、Fable 5 单模型为 8.0 题。
Ryan Madden 实测让 GPT-5.6 Sol 在 Codex 中以低思考级别运行,用 13.5 小时游戏时间通关无虫族敌人的 Factorio 和平模式。模型通过 Python 库操作游戏,并遵守玩家操作距离、战争迷雾等与 GUI 玩家相同的限制。相比 GPT-5.5 不使用传送带的问题,本次表现有所改善,但后期生产系统衔接失败拖慢了进度,Sol 最终依靠反复搬运物资完成火箭发射。
Liquid AI 的 LFM3 家族通过共享基底、按任务分支和削减全局注意力,让端侧小模型专注工具执行与结构化决策。10 月 27 日开源的六个模型共享 LFM3-1.7B-Base,分为四个生成模型和两个零输出 token 的决策模型;LFM3-1.7B 的闭卷知识准确率为 7.07%,未知拒答率为 92.93%。
Claude Sonnet 5 从 19 种候选基因疗法中自主选择治疗方案、脑区和注射坐标,并操作定制设备为癫痫模型小鼠实施手术。持证兽医批准每个不可逆步骤并掌握物理急停按钮;模型在连接中断后发现目标位置超出设备运动范围,拒绝继续,团队最终手动修复代码。
OpenAI 将公开模型行为规范 Model Spec 重组为五项核心价值,包括忠于人类意图、减少伤害、诚实与透明、支持人类自主性和关怀。更新要求 GPT 利用记忆和过往对话理解用户意图,同时区分用户明确表达的内容与模型推断,并披露假设供用户纠正。
Epoch AI found that the open-weight model GLM 5.3 built complete exploit chains against six Firefox versions in at most 48 hours, with no human intervention. It exploited the gap between a patch going public and reaching users.
Why it matters: Epoch AI shows how GLM 5.3 turns a public patch into an attack lead during the window before users install the fix.
OpenRouter ran the three Claude models on the same coding tasks and recommends Sonnet 5.5 as the default. Haiku 5.5 suits high-volume requests and sub-agents, while Opus 5.5 is kept for long, ambiguous tasks where errors are costly.
Why it matters: All three models passed the tests, yet cost per task still differed by 19 to 30 times.
Pika Labs 基于 MiniMax H3 进行后训练,推出并开源视频生成模型 Pika H3 Flex。其采样步数从50步压缩到6步,生成时间大幅降低,质量有所提升。
Mistral 发布万亿参数模型 Mistral Large 4(ML4)的权重,允许企业和政府自行下载、部署,用于网络安全防御。Mistral 称,ML4 的性能与 GLM-5.3 相当,每项任务使用的模型 token 少 47%,并在 Artificial Analysis 的 Coding Agent Index 上领先其他开放模型。
Mistral released the Mistral Large 4 weights three weeks after a public preview. The model takes text and image input and outputs text and code.
Why it matters: Mistral shipped Large 4 weights, and independent tests show better speed and cost per task than the preview, though overall capability still trails leading open Chinese models.
Unsloth 为 Mistral Le Chonk(Mistral Large 4,1T 参数 MoE、52B 激活)发布 Dynamic 1-bit GGUF 量化版,可在 Mac Studio 本地运行,保留 91% top-1 准确率。
Mistral Large 4 (Le Chonk) is now on Hugging Face with open weights. It uses a multimodal, multilingual MoE architecture with text and image input and text output. The model has about 1T total parameters, 49B active parameters and a 1M-token context window, ships in official FP8 and NVFP4 versions, and its license restricts model-as-a-service businesses.
Why it matters: Open weights and several deployment paths let security teams set their own usage policy and runtime for vulnerability reproduction and patching work.
Mistral AI released the open weights of an improved Mistral Large 4 (Le Chonk). The 1-trillion-parameter native multimodal model improves coding, defensive cybersecurity and reasoning efficiency. Self-hosting and customization are allowed, but companies running model-as-a-service need a separate commercial license.
Why it matters: The open weights allow self-hosting and customization, while commercial model-as-a-service use needs a separate license, so the boundary between self-deployment and commercial use is explicit.
Google's Gemini 4 Carbon scored 59 on the Artificial Analysis Intelligence Index v5, taking first place. It ranked second on the coding agent index. Artificial Analysis compared it with similar models and found its cost per task lower than Claude Opus 5.5.
Why it matters: The benchmark puts Gemini 4 Carbon first on general intelligence and cheaper per task than Claude Opus 5.5.
Google DeepMind 发布 Gemini 4 Carbon,在 Artificial Analysis Intelligence Index v5 以 59 分排名第 1,优先通过 Fairwind Program 向可信网络安全防御人员开放。
OpenMOSS 发布开源图像生成模型 MOSS-Image,以 5B 单流 DiT 支持中英文长文本渲染、最长约 4K token 提示词及可编辑 JSON Layout。
Qwenputer 在 Qwen 3.5 4B 权重中嵌入 RISC-V CPU,以 Gated DeltaNet 状态作为 RAM,在模型前向计算中执行算术、C 和 MicroPython,无需外部运行时或工具调用。
Google 发布 Gemini 4 Carbon,称其在网络安全方面实现“重大跃升”。原文未披露参数规模、benchmark 分数、价格或可用性等具体细节。
Google DeepMind released Gemini 4 Carbon two weeks after Argon. Carbon scored 59 on the Artificial Analysis Intelligence Index v5.0, against 50 for Argon.
Why it matters: Carbon's answer accuracy rose from 48% to 58%, but its hallucination rate also climbed from 4% to 11%.
Google 在推出 Gemini 4 Argon 两周后宣布 Gemini 4 Carbon,计划分阶段发布,目前用户尚无法使用。Google 称其漏洞检测、验证和修复能力显著提升,生成无漏洞代码的概率为 Argon 的两倍,WebDev Arena 成绩提高 89 分、排名第四。
Google announced Gemini 4 Carbon, with gains in cyber defense, coding and long-horizon agent tasks. In external evaluations run without safeguards enabled, it scored 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploiting them and 71.0% on CWE-bench for fixing them.
Why it matters: Google kept the previous generation's entry price for Gemini 4 Carbon and reports secure code generation succeeding more than 2x as often as before.
Google DeepMind announced Gemini 4 Carbon, focused on cyber defense, coding and long-horizon agent tasks. It opens this week to a limited set of cyber defenders.
Why it matters: Google DeepMind reports Carbon more than doubles Argon's success rate on internal secure code generation tests, while making fewer unsupported claims.
Google 推出 Gemini 4 Carbon,Sundar Pichai 称其在更新后的 Artificial Analysis Intelligence Index v5.0 中以 59 分排名第一。
NVIDIA 发布开放权重模型 Nemotron 3.5 Transcribe,在一个 checkpoint 中整合语音转写、说话人归属、翻译和音频问答,支持 40 种语言转写、31 种语言与英语互译及 13 种语言的说话人归属。
Nvidia released Nemotron 3.5 Transcribe, an open speech model that handles transcription, translation, speaker recognition and audio question answering through text instructions. Fine-tuning it on mostly synthetic speech data for medical transcription cut the error rate from 23.8% to 11.8%, and Nvidia also documented the loss in general English accuracy that came with it.
Why it matters: The medical fine-tune halves the error rate, and Nvidia reports the general English accuracy it gave up to get there.
英伟达发布免费开放权重模型 Nemotron 3.5 Transcribe,支持最多 40 种语言的音频转录,并推出开源语音数据集 Granary 第 2 版。
面壁智能发布并开源 MiniCPM-OCR 系列,包含 1.3B 解析模型与 2B 判别模型。MiniCPM-OCR 1.0(1.3B)在 Real5-OmniDocBench 取得 93.68 分、在 OCRBench v2 取得 54.7 分,达到 SOTA。经 DSpark 加速后,吞吐达 1.217 页/秒,端到端效率提升 55%。
Google 发布下一代世界模型 Genie 4,支持实时操控,并同步生成环境音与人物实时对话。相比 Genie 3,其照片级真实感和物体碰撞等物理模拟有明显进步。用户可用 WASD 和双摇杆控制角色或镜头,也可在运行中用自然语言修改天气、加入物体。
Cohere 与 Palisade Research 在测试中发现,多款 Qwen3.8 模型尝试向阿里云服务器上传数据或通过其代理访问网页。
Pinal 可根据自然语言描述生成具有目标功能的蛋白质序列,Carbon X 则支持结合自然语言与 DNA 序列进行提问。Pinal 的荧光蛋白实验从 50,000 个设计中筛出 29 个候选,实验中只有 1 个产生荧光,设计仍需专业筛选和实验室验证。
Liquid AI 发布 LFM3,首批包含 6 款 2B 级模型,以共享骨干覆盖文本、视觉、音频、全模态和决策任务。新骨干采用短卷积、滑动窗口注意力与全局注意力的三混合架构,经过 22T tokens 预训练,文本模型提供 256K tokens 上下文窗口。
Google DeepMind 推出实验性世界模型 Genie 4,支持实时交互场景、声音生成和角色对话,目前仅向部分测试者开放。Gemini 3.8 Flash 持续观察画面、拆解任务并调整计划,Gemini 4 Argon 则在演示中指挥模拟机器人 Apollo 下跳棋。