跳到正文
8月30日周日
  1. Simon Willison 博客66

    腾讯发布 Hy4 Preview:770B 参数开源权重文本模型,上下文窗口达 1M token

    腾讯今日发布 Hy4 Preview,这是一款新的开源权重纯文本输入(无视觉)大语言模型,总参数 770B,激活参数 49B,上下文窗口 1M token,Hugging Face 上体积 1.56TB。相比 7 月的 Hy3(295B 总参数、21B 激活、256k 上下文、598GB),规模大幅提升。Hy4 的推理配置仅支持 "high"(默认)和 "no_think" 两档。

  2. Rohan Paul50

    OpenRouter 周 token 量从 2024 年 12 月至 2025 年底增长 25 倍,2026 年初又翻三倍,呈指数级复合增长。Cathie Wood 指出,token 推理量一年增 25 倍并持续倍增,可能推动实际 GDP 年增速迈向两位数,但多数客户与公司仍难以理解这一变化。

    引用Cathie Wood@CathieDWood

    When I try and explain that tokens inferenced increased 25-fold in a year and are multiplying exponentially, cascading through the economy, and are likely to push real GDP growth per year toward double digit territory, many clients and companies simply do not comprehend.

  3. Tencent Hy52

    腾讯混元开源 Hy4 preview,旗舰 MoE 模型,支持 1M 上下文,采用 Apache 2.0 协议。模型总参数量 770B,每 token 激活 49B,原生集成 MTP 层支持投机解码。163 位内部专家在 203 项真实工程任务中的盲测中,Hy4 preview 平均分 2.99,优于 GLM 5.3(2.92)和 Kimi K3(2.94)。

    引用ModelScope@ModelScope2022

    Tencent HY open-sourced Hy4 preview: a flagship MoE with 1M context, and their biggest generational leap measured to date. Apache 2.0.🔥 🤖 https://modelscope.cn/models/Tencent-Hunyuan/Hy4-preview ⚙️ 770B total params, 49B active per token. Ships with a native MTP layer for speculative decoding out of the box. 📊 Blind eval by 163 internal experts across 203 real engineering tasks: it edges out GLM 5.3 (avg 2.99 vs 2.92) and Kimi K3 (2.99 vs 2.94). 💼 Trained on real production workflows: software engineering, office and data analysis, game development, and scientific research. 🌍 For global users: https://modelscope.ai/models/Tencent-Hunyuan/Hy4-preview

8月29日周六
  1. Rohan Paul43

    EdgeBench 长时程基准测试中,最强智能体运行 12 小时仅得 51.3/100 分。该基准提供 134 个任务的持久执行环境与反馈,智能体可调试代码、运行模拟并反复提交结果,但 6 类任务约 3.8 万小时交互仍难转化为持续进步。5 个模型的平均表现呈 log-sigmoid 曲线(R²≥0.997),先慢后快再饱和。

    引用Rohan Paul@rohanpaul_ai

    "Long-horizon tasks are still a joke. They do not work, and I do not care what anybody says. Do not show me a stupid evaluation. Do not tell me about some dumb script you ran for 48 hours. Long-horizon tasks are not handled well. They simply do not work." - Chamath at Stanford AI Club "2nd, complex problems also do not work. They are neither addressed nor handled well. Why is this important? If AI develops like any other technology, we are going to experience an initial rise—the hype cycle. Then, we will see a natural contraction because, somehow and somewhere, something is going to fail. We are all going to see this, and then we will enter what is called the “trough of disillusionment.” I think the business and MBA folks will confirm whether that is true. Afterward, you typically see the slow and gradual adoption of the real, final solution. This happened with the internet, and it has happened in many other cases. The problem is that we are spending hundreds of billions, potentially trillions, of dollars trying to figure out how to cross this chasm. So, what do we do? If we do not figure this out, people will reach the trough of disillusionment and say that AI was a joke. I think we need to be able to bring AI into highly complicated environments and make it work. What is my solution? At a very basic level, you need a symbolic space that guides the embedded space." ---- From "techniahqrobot" YouTube channel, (full video link in comment)

  2. IT之家·人工智能46

    行业首个量子增强型大模型“玄幂 Xenomi”发布:深度融合 AI 与量子计算

    国光量超人工智能研究团队推出行业首个量子增强型大模型玄幂(Xenomi)家族,将量子技术引入大语言模型的决策与推理过程。该模型面向科研和学术领域,在科研科普、路由和智能体决策等任务中较主流开源模型展现出更稳定的专业理解、更清晰的判断和更短的决策路径。玄幂已应用于国光量超的量子计算智能实验平台和夸米多智能体系统。

  3. Rohan Paul42

    Long-Horizon-Terminal-Bench 基准测试显示,17 个前沿模型在 46 个长程终端任务上平均通过率仅 6.4%,严格全完成评分下 10 个模型零通过。失败运行中 79% 因超时终止,最佳模型也仅达 28.3%。模型能执行局部合理步骤,却无法将数百步转化为最终成果,印证 Chamath 关于长程任务“根本行不通”的论断。

    引用Rohan Paul@rohanpaul_ai

    "Long-horizon tasks are still a joke. They do not work, and I do not care what anybody says. Do not show me a stupid evaluation. Do not tell me about some dumb script you ran for 48 hours. Long-horizon tasks are not handled well. They simply do not work." - Chamath at Stanford AI Club "2nd, complex problems also do not work. They are neither addressed nor handled well. Why is this important? If AI develops like any other technology, we are going to experience an initial rise—the hype cycle. Then, we will see a natural contraction because, somehow and somewhere, something is going to fail. We are all going to see this, and then we will enter what is called the “trough of disillusionment.” I think the business and MBA folks will confirm whether that is true. Afterward, you typically see the slow and gradual adoption of the real, final solution. This happened with the internet, and it has happened in many other cases. The problem is that we are spending hundreds of billions, potentially trillions, of dollars trying to figure out how to cross this chasm. So, what do we do? If we do not figure this out, people will reach the trough of disillusionment and say that AI was a joke. I think we need to be able to bring AI into highly complicated environments and make it work. What is my solution? At a very basic level, you need a symbolic space that guides the embedded space." ---- From "techniahqrobot" YouTube channel, (full video link in comment)

  4. Noam Brown41

    OpenAI 研究员 Noam Brown 回应为何未公布 Astra 更多数学成果:内部版 Astra(OpenAI 下一代模型家族)已解决数学、量子复杂性与理论计算机科学领域的 10 个重大开放问题。OpenAI 仅在成果能显著改变人们对 AI 进展速度认知时才公布,当前重心是发布优质模型供大众使用。

    引用Noam Brown@polynoamial

    An internal version of Astra, @OpenAI’s next major model family, solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science. We believe it will be a major step for scientific reasoning. https://openai.com/index/ten-advances-in-mathematics/

  5. Apple Machine Learning Research52

    LLM 并非(始终)符合贝叶斯:量化 LLM 概率信念的内部(不)一致性

    苹果机器学习研究团队提出一种新方法,将 LLM 视为信息处理规则,利用其与贝叶斯更新的信息处理差距,研究模型如何根据证据更新概率信念的内部(不)一致性。实验评估了 LLM 在医学、科学、法律等复杂领域的信念更新表现,揭示其概率推理与贝叶斯理想之间的系统性偏差。

  6. OpenRouter52

    来自 @Zai_org 的 GLM-5.3 现已开放权重,并上线 OpenRouter。 该模型专为复杂软件工程、长周期智能体和网络安全设计,支持 1M 上下文和可配置的推理强度。 立即体验:https://openrouter.ai/z-ai/glm-5.3

    引用Z.ai@Zai_org

    GLM-5.3 现已开放权重。 我们最强大的智能体编码与网络防御模型,现已开放下载、运行和定制。 权重:https://huggingface.co/zai-org/GLM-5.3 技术博客:https://z.ai/blog/glm-5.3

8月28日周五
  1. Rohan Paul44

    Sapient Intelligence 发布并开源自主科研系统 PRAXIST Beta,通过并行 Research Peers 与共享研究图谱跨代累积实验证据。在完整 75 任务 MLE-bench 上,PRAXIST(deepseek-v4-pro)获 49 金、共 60 枚奖牌,超过 Claude Opus 4.8 的 34 金、55 枚。

    引用Sapient Intelligence@Sapient_Int

    Introducing PRAXIST Beta, your autonomous research team🚀 Define the objective, constraints, and what success looks like. PRAXIST discovers the path. From there, PRAXIST takes on the experimental research loop. Multiple Research Peers explore competing approaches in parallel, share useful findings, and build on accumulated evidence across experiments and generations. In partner-provided environments, PRAXIST reached a 100% safe-landing rate in a rocket simulation and reduced accumulated error in an industrial SLAM system from 9.37 centimeters to 5.01 centimeters. From robotics control to quantitative finance, PRAXIST has delivered measurable advances across fundamentally different problems by combining a shared core research architecture with domain-specific tools, knowledge, constraints, and evaluators. More approaches explored. Faster iteration. Greater R&D capacity without proportionally increasing specialist headcount.

  2. Anthropic:Alignment Science Blog59

    Anthropic 发布 CHIVE 管线:用反事实实验评估 LLM 行为解释,发现可解释性工具未带来提升

    Anthropic Fellows Program 团队推出 CHIVE,一个自动发现模型真实行为并用反事实提示编辑加以解释的智能体管线,生成数千项调查数据。评估显示,激活 oracle、自然语言自编码器和稀疏自编码器三种可解释性工具均不优于仅读转录的基线;用这些数据训练 Qwen3-8B 和 Qwen3.5-397B-A17B 预测反事实结果,可泛化到未参与训练的设置。论文与代码已开放。

8月27日周四
  1. The Decoder:AI News41

    OpenAI 研究员警告:超快 AI 推理速度或让安全团队措手不及

    一位化名“roon”的OpenAI研究员警告,超快AI推理带来现有防护无法应对的安全风险。一个未对齐的模型若以当前最佳系统50倍的速度运行,可能迅速渗透系统,人类响应团队将难以招架。roon认为仅靠监控不够,需要自主检测与关停机制;该评论针对OpenAI新AI芯片发布,其推理速度优于现有硬件。

  2. The Decoder:AI News69

    GLM-5.3-Flash 发布:3200 亿参数、百万上下文,成本仅为 GLM-5.3 的约 1/7.5

    Z.ai 发布 GLM-5.3-Flash,3200 亿总参数(180 亿激活)、百万 token 上下文窗口,MIT 许可开源。该模型在 Intelligence Index 上得 57 分,接近 GLM-5.3 的 60 分,但每任务成本仅 0.09 美元,约为后者的 7.5 分之一。该模型全程运行于国产 AI 芯片,Z.ai 自研服务软件实现了与 Nvidia GPU 相当的性能。

  3. IT之家·人工智能72

    我国日均词元调用量突破 500 万亿,中国大模型稳居全球第一梯队

    截至 2026 年 6 月,我国日均词元调用量已突破 500 万亿,中国大模型在全球竞争中位居第一梯队。当前旗舰模型几乎以月为单位更新,竞争焦点转向智能体落地与生态建设,推理算力需求随之爆发。腾讯混元 3 正式版上线第一周,Token 调用量比上一代混元 2 增长 68 倍。

    推荐理由:模型更新周期缩短至4到6周,且智能体任务的多轮调用说明推理算力需求上升,对算力产业链的判断需从训练侧扩展到推理侧。

  4. IT之家·人工智能50

    小鹏刘先明:第二代 VLA 升级核心是让模型首次建立对时间维度的理解

    小鹏集团通用智能中心负责人刘先明在物理 AI 分享会上表示,第二代 VLA 升级的核心是让模型第一次建立起对时间维度的理解。小鹏物理世界基座模型首次将“时间”纳入模型,AI 理解维度从“3D 空间”跃迁至“4D 时空”,并引入 Infini-VLA 长时序架构,可记住前 30 秒的世界。第二代 VLA 同步采用流式自回归推理,端到端响应提速 300%,实现“边看、边想、边行动”的并行处理。

  5. Qwen50

    Qwen3.8-Flash 现已上线 @OpenRouter!⚡ 编程助手、智能体工作流、长视频理解,一个 API 调用即可全部实现。

    引用OpenRouter@OpenRouter

    Qwen3.8 Flash from @Alibaba_Qwen is now live on OpenRouter. A multimodal reasoning model for coding assistants, agentic workflows, visual understanding, codebase and document analysis, desktop interaction, charts, and long video. https://openrouter.ai/qwen/qwen3.8-flash

  6. 公众号:龙猫LongCat57

    GeoRA:为RLVR设计的LoRA——ACL 2026杰出论文解析

    美团与北京大学提出GeoRA,将低秩适配显式对齐到RLVR的更新几何,先用谱先验与欧氏先验定位稀疏更新区域,再用SVD压缩为低秩稠密适配器并冻结残差锚点。在1.5B至32B的Qwen与Llama模型上,GeoRA在数学、医学、代码三类RLVR任务上稳定优于LoRA、PiSSA、MiLoRA等基线,Qwen3-8B的AIME24达23.75%,可训练参数较全参微调降低99.5%。

  7. IT之家·人工智能48

    黄仁勋回应 OpenAI 首款自研 AI 芯片挑战:英伟达提供“截然不同”服务

    黄仁勋在财报电话会议上回应 OpenAI 自研推理芯片 Jalapeño,称英伟达提供的是覆盖 AI 生命周期的完整 AI 工厂平台,可在任何云环境使用,与针对单一模型的定制化 ASIC 芯片有本质区别。他 100% 确信英伟达技术对 OpenAI 仍将卓越。此前测试显示,Jalapeño 在 DeepSeek R1 上每瓦 AI 吞吐量是 GB300 的 1.7 倍。

  8. IT之家·人工智能55

    星辰技术二期中试线9月通线,主攻3D先进封装,华为韬定律加速落地

    湖北星辰技术有限公司完成近50亿元A轮融资,刷新国内先进封装领域融资纪录,规划月产10万片。其二期中试线预计9月全线通线,一期二期合计投资超70亿元,月产能规划2万片,已建成国内规模最大高密度集成封装中试和小批量量产平台。通过先进封装,星辰技术将推理带宽比传统结构提升1到2个数量级,正协助国内头部企业打磨三维计算芯片。

  9. IT之家·人工智能74

    英伟达 2027 财年半年报归母净利润 1180.1 亿美元,同比增长 161.1%

    英伟达发布 2027 财年半年报,上半年营收 1778.37 亿美元,归母净利润 1180.1 亿美元,同比增长 161.1%,GAAP 毛利率 75%。第二财季营收 962.21 亿美元,同比增长 106%,环比增长 18%,归母净利润 596.88 亿美元,同比增长 126%。数据中心业务第二季度收入 890.23 亿美元,同比增长 117%,Vera Rubin 平台已进入全面量产。

    推荐理由:数据中心收入同比翻倍叠加 Vera Rubin 平台全面量产,显示 AI 基础设施的资本开支仍在加速,延迟部署的概率下降。

8月26日周三
  1. ByteByteGo47

    如何让 LLM 提速 3 倍:投机解码原理详解

    投机解码通过让一个小模型预先草拟多个候选 token,再由大模型在单次前向传播中并行验证,可将生成速度提升 2-3 倍,且输出在统计上与单独运行大模型完全一致。70B 参数模型每生成一个 token 需从 GPU 内存读取约 140 GB 权重,而生成阶段计算利用率仅 20-40%,投机解码正是利用这部分闲置算力。文章还探讨了草稿来源的四种方式及投机解码失效的场景。