Google announces Gemini 4 Carbon, its latest frontier AI model that you can't use
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 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。
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。
Google DeepMind 发布 Genie 4 世界模型,称其相比 Genie 3 有多项改进,并承诺更逼真的物理模拟与人物。世界模型根据用户描述,在探索过程中逐帧实时生成体验,无需预渲染环境或编写交互元素。DeepMind 称该技术可用于物理机器人的训练环境和创意领域的快速原型制作,Genie 4 的改进部分得益于更新后的 Gemini 模型。
Google 今天宣布世界模型 Genie 4,可生成更逼真的交互体验,支持用自然语言提示词实时修改体验并生成音频。模型使用 Gemini 3.8 Flash 辅助模拟,体验时长从 Genie 3 的一分钟增至约两分钟,Google 提醒较长模拟中的细节和物体持续性可能发生变化。
Google DeepMind 于周二发布 Genie 4 研究预览,新增实时音频与模拟中的提示词编辑,生成更逼真、可交互的世界。模型可实时改变天气、镜头角度或添加人物与物体,保持场景一致性最长两分钟,是前代的两倍,但仍存在细节漂移、音频准确性和场景文字可读性问题。
Google DeepMind released Genie 4, a world model with native audio, interactive dialogue and real-time scene edits in natural language. It is in research preview for trusted testers. In simulations, Gemini 3.8 Flash keeps observing the scene, breaking down goals and checking progress to hold multi-step events together. The model still drifts on object persistence, complex physical interactions and long-term consistency.
Why it matters: Genie 4 puts native audio and live scene control into world simulation, so creators can adjust camera, dialogue and action to try out ideas in a moving scene.
OpenAI 发布 GPT-6.1 Terra,强调细致的视觉推理、精准的计算机操作,以及更清晰自然的写作。OpenAI 的 Ted Sanders 表示,Terra 的技术沟通可能优于 Astra 和 Sol,视觉与计算机操作能力相近且速度更快,多数其他能力略弱。
OpenAI released GPT-6.1 Terra, cutting API prices and improving results on reasoning, coding and computer-control benchmarks. Input and output cost $1 and $5 per million tokens, down from $2 and $12 for GPT-5.6 Terra. Artificial Analysis measured a 6-point intelligence index gain at its highest reasoning effort, with cost per task falling from $1.40 to $0.40.
Why it matters: Independent testing puts GPT-6.1 Terra 6 index points above its predecessor, with cost per task down from $1.40 to $0.40.
OpenAI released GPT-6.1 Terra, which combines visual reasoning, computer operation and natural writing. Standard API pricing is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol.
Why it matters: OpenAI cut the standard API price of GPT-6.1 Terra by 50%, and its computer-use benchmark score is 15 percentage points higher than GPT-5.6 Terra.
GPT-6.1 Terra 可将产品照片匹配到目录条目、解读技术图纸、操作应用屏幕,以及起草和修改文本。reasoning.effort 支持 none、low、medium(默认)、high、xhigh 和 max。工具调用要求使用 Responses API,提示词与迁移指导见 GPT-6 指南。
OpenAI has released GPT-6.1 Terra, pitched on writing, personality, close reading of requests and computer use. The release post calls it the company's first model that is clearly better at writing, personality and fine-grained request understanding, and says it can read what is on screen and act across apps. It offers twice the usage of 6.1 Sol in the API and subscriptions, at a lower price.
Why it matters: GPT-6.1 Terra gives twice the usage of 6.1 Sol in the API and subscriptions, and costs less.
OpenRouter 宣布 OpenAI 的 GPT-6.1 Terra 上线,定价为 $1/M input、$5/M output,上下文为 1.05M。该模型在 Agents' Last Exam 上与 GPT-6 Astra 得分相当,每 token 价格约为其十分之一,平均完成任务速度比 Astra 和 GPT-6.1 Sol 快 3 倍以上。
OpenAI released GPT-6.1 Terra, rolling it out in ChatGPT Work and Codex to Plus, Pro, Enterprise and Edu users starting today. The model is pitched at speed and efficiency, with stronger visual reasoning, precise computer control and clear, natural writing. OpenAI lists reading technical drawings, drafting proposals, filling in forms and editing video among its work uses.
Why it matters: The model ships straight into ChatGPT Work and Codex, so existing paid users can test it on drawing review, forms and video editing.
Unsloth 发布 MiniMax-M3.1-Flash 的量化 GGUF,2-bit 版体积缩小 45% 的同时保留 91.5% top 1 精度,可在 150GB 内存上本地运行。1-bit 版精度约 89.9%,4-bit UD-Q4_K_XL 达 98.0%。作者称其为同尺寸最强开源模型之一,并给出 Hugging Face 权重链接与本地运行指南。
腾讯混元于 10 月 26 日发布并开源同声传译大模型 Hy-SimulMT,支持中英语音到语音实时翻译。实测响应时间约 3 秒,低于人类专业译员平均的 4-5 秒。
Destined AI 发布开源语音合成模型 Hurston-1,基于 Maya1 微调,支持美国南部及多个非洲地区的英语口音。模型通过自然语言描述控制声音,默认输出 24 kHz 音频,电话模式输出 8 kHz G.711 音频以接入 Twilio 等电话系统,模型权重采用 Apache-2.0 许可。
南京大学、商汤及其他高校的研究团队发布 OpenMobile-2 两款手机智能体模型,分别基于 Qwen3.5 和 Qwen3.6,参数量为 9 billion 和 27 billion,并开放训练环境、数据及复现说明。
Dream Labs 发布世界动作模型 Dream-1 并开放公共 API,称其可零样本控制未见过的双臂平行夹爪机器人,执行未见任务。评估覆盖 27 种模拟机器人和 4 种真实机器人,表现因机器人及其逆运动学求解器而异;FaaS 微调服务可利用机器人自身数据改善适配。
Block 发布可本地运行的全双工语音模型 Dessa,支持同时听说,并展示了笑声、停顿和哭声等接近真人的表达。Block 称该模型完全使用人类语音和单张 NVIDIA GPU 训练,知识来自对话数据,情绪表达内置于模型。Dessa 可在 MacBook M4 Max 或更高配置的消费级电脑上运行,目前尚未公开开放,Block 正邀请安全研究人员测试。
Block 发布 AI 语音模型 Dessa,号称首个完全基于人类语音、不含任何文本数据从零训练的模型,直接以音频处理语音,可实时捕捉情绪、讽刺、笑声和停顿。该模型支持全双工交流,可在 MacBook Pro M4 Max 本地运行,训练使用单台 NVIDIA GB200 NVL72。Dessa 尚未公开,团队正招募 AI 安全研究者测试其意外风险。
Moonshot AI released Kimi K3.1 Pro, available now through the API, with weights promised for November 4, 2026. The MoE model has 2.8T total parameters and 104B active, and a 1-million-token context window. Moonshot says the Highspeed version keeps the same intelligence while outputting 2–3× faster, at twice the standard price.
Why it matters: The per-token price is a third lower than K3, but independent tests show per-task cost flat or higher because the model uses more tokens.
Moonshot AI released its flagship model Kimi K3.1 Pro, with 2.8T parameters and a 1M-token context window. Full model weights are expected to open up by November 4. API pricing is 13 yuan per million input tokens and 68 yuan per million output tokens, roughly half of K3. A Highspeed version outputs about 2 to 3 times faster at double the price.
Why it matters: Moonshot AI roughly halved Kimi K3.1 Pro's API price against K3, while the Highspeed version trades double the price for about 2 to 3 times the output speed.
Block 旗下 dessa.com 推出对话模型 Dessa,仅使用语音数据训练。该页面当前展示的是 Block 于 2026 年 10 月 21 日生效的隐私声明,说明其收集的姓名、邮箱、设备信息及使用数据,并列出与关联公司和第三方服务商共享信息的范围。
Jack Dorsey 的 Block 宣布语音模型 Dessa,仅用人类对话从零训练,演示交流大体自然,也出现认错人和突然跑题。团队称它是直接处理语音输入与输出的模型,所有知识均来自人类语音,可在 MacBook Pro M4 Max 或更高配置设备上本地运行。团队希望改善人与 AI 的理解及语音情绪识别相关安全问题,将向一小群 AI 安全研究人员提供早期访问权限。
Kimi K3.1 Pro went live on the Kimi API today. The weights are not open yet; Moonshot AI says it will release the full model weights on Nov 4, 2026. Its input and output pricing match GPT-6.1 Sol and come in at half of Claude Opus 5.5.
Why it matters: Kimi K3.1 Pro prices input and output at the same rate as GPT-6.1 Sol, half of Claude Opus 5.5.
Moonshot AI released the reasoning model Kimi K3.1 Pro. It scores 55 on the Artificial Analysis Intelligence Index, ranking fourth, 11 points above Kimi K3 and 3 points behind Claude Opus 5.5.
Why it matters: Kimi K3.1 Pro adds 11 index points at the same model scale, while cost per test task drops from $2.00 to $1.98.
Moonshot AI released the reasoning model Kimi K3.1 Pro, which scored 55 on the Artificial Analysis Intelligence Index to place 4th, an 11-point gain over Kimi K3.
Why it matters: Kimi K3.1 Pro improves sharply over Kimi K3 on agent benchmarks while cost per task stays around $1.98.
Moonshot AI released Kimi K3.1 Pro, which supports a 1M token context, and says the full model weights will be released by November 4. A companion model, Kimi K3.1 Pro Highspeed, runs at 2-3x the speed. Both are available from today on Kimi.com, Kimi Work, Kimi Code and the Kimi API.
Why it matters: Moonshot disclosed vulnerability-task results before and after removing cybersecurity training data, giving a concrete comparison for risk control in open weights.
Kimi made K3.1 Pro generally available on Kimi Code and released K3.1 Pro Highspeed, which has the same capabilities and outputs about 2–3× faster. K3.1 Pro has 3T parameters and a 1M context window, accepts image and video input, and runs only in thinking mode. Kimi says coding, 3D creation and financial research and modeling are notably improved.
Why it matters: The Highspeed variant trades about 2× the quota for roughly 2–3× faster output, while tool calls and script execution do not speed up.
Kimi released K3.1 Pro, a 3T-class model with a 1-million-token context window, and K3.1 Pro Highspeed, which runs 2–3× faster. After Kimi added cybersecurity risk controls to training data, K3.1 Pro's pass rate on an internal exploit test fell from 90.5% to 81.0%.
Why it matters: The drop from 90.5% to 81.0% on an internal exploit test shows what safety filtering in training data costs in offensive-security capability.
月之暗面发布 Kimi K3.1 Pro,称其为自家最强的 3T 级开放权重模型。该模型采用 MoE 架构,总参数为 2.8T、激活参数为 104B,支持 1M token 上下文与多模态。