Chinese contentElevenLabs 发布 Eleven v4，八语言盲测中听众认为其表现力更强
ElevenLabs 放出了自家 Eleven v4 和竞品的八语言盲听测试，听的人都说它读起来更有感情，做配音的可以去听听对比。
Latest
Every published item, in order · 1029 items
Oct 16
Oct 15
Chinese contentOpenCode 向用户免费开放 DeepSeek V4.1 Flash
用 OpenCode 写代码的可以白嫖 DeepSeek V4.1 Flash 了，免费 200K 上下文，要 1M 才需要付费。
Chinese contentDeepSeek 服务恢复，疑似新模型上线，生成速度约 180 t/s
有人实测 DeepSeek 服务恢复后的新模型，速度 180 t/s，表现接近 Claude，可能是 4.1 Pro，想去试的可以看看。
Chinese content研究拆解 LLM 网络威胁情报实体抽取失败模式：主要问题不是幻觉
DeepSeek-V4-Flash 和 GPT-6.1-Sol 抽实体都不靠谱，但坏的原因不是幻觉，而是把背景信息当成攻击实体。用置信度分层排审核优先级，别当真值。
DeepSeek V4.1 Flash Tops Artificial Analysis AgentPerf Leaderboard
Serving efficiency, not raw scale, is where V4.1 Flash wins: it delivers 1.3x to 6.2x the throughput per MW of V4 Pro 0813 across Blackwell hardware while scoring higher on intelligence. Teams running agentic coding workloads can use the AA-AgentPerf data to pick hardware — B300 nodes stay within about 20% of rack-scale GB300 NVL72, so single 8-GPU nodes are viable for this model. As an open-weights release from DeepSeek, it widens cheap self-hosting options for Southeast Asian inference providers competing with rented frontier APIs.
AMD MI355X joins AA-AgentPerf benchmark against NVIDIA Blackwell
The MI355X's 288GB of HBM3E per GPU gives it an edge on very large models like the 1.6T-parameter DeepSeek V4 Pro, but serving-stack maturity on ROCm remains the bottleneck at higher loads. Teams evaluating accelerator purchases for agentic inference can now compare AMD and NVIDIA on realistic coding-agent workloads rather than synthetic throughput numbers. All results run on ROCm nightly builds of vLLM and SGLang, so scores should improve as the software stack matures.
BizBench Benchmark Tests Agent Execution Validity in Payments Workflows
The benchmark shows that outcome-only evaluation masks invalid execution, since models can hit the final goal while violating constraints mid-run — a gap that matters most in regulated domains like payments where intermediate errors carry financial or compliance consequences. Agent builders can use the 50-task suite and deterministic checks to measure constraint compliance and tool-call overhead, not just task success. Teams choosing models for workflow automation get a direct comparison of frontier and open models under identical conditions.
Chinese contentAtomSearch：用证据驱动的树搜索优化 Agent 技能
一篇讲怎么高效改 Agent 技能规则的论文，AtomSearch 在 deepseek-v4-flash 上比 SkillOpt 高 9.68 个百分点，技能还短好几倍。
Chinese content论文提出 Memory Echo 现象与 FPA 框架，修复智能体长期记忆失真
做智能体长期记忆的可以看看：论文发现同源旧记录会重复刷存在感，FPA 框架实测把准确率拉高近 10 个点，代码开源可复现。
Chinese contentarXiv 论文：证伪驱动的模型无关智能体协议，输出纪律决定决策下限
用中医辨证基准证明了件事：很多模型翻车不是能力不够，是输出纪律不行。他们设计的证伪协议能把 Kimi K3 的方差从 0.218 压到 0.007，写智能体的值得看看。
Chinese contentANIS 框架为智能体生成原生桌面界面，成功率达 74.5%
给智能体设计界面这事有论文了，DeskIF-Bench 和 ANIS 都开源思路，做 agent 桌面自动化的可以看看怎么提分还省 73% token。
Chinese contentSeTPO 训练方法让小语言模型更擅长调用外部工具
小模型学用工具的老难题有了新解法：SeTPO 用双向 KL 同时管住探索和模仿，BFCL v4 上跑赢了蒸馏和纯 RL，做端侧模型的可以看看。
Chinese contentAI智能体删掉The Challenge整季文件？电视台归咎备份失效
制作公司让AI智能体接入服务器，整季节目剪辑文件没了，首播拖到明年，电视台坚持说是备份问题，你信谁。
IT之家Source
Chinese contentDeepSeek V4.1 Flash发布开源演示应用
DeepSeek用V4.1 Flash做了个开源演示应用，在Together AI上能直接体验，感兴趣的朋友可以去试试。
Oct 14
Chinese contentBRQGen：迭代式评估 Bug 报告完整性并按需生成追问
这篇提出了一个挺聪明的思路：不再机械地问固定数量的问题，而是按报告缺口按需追问，作者还放了 429 条真实数据集和多个模型的对比成绩。
Chinese contentDeepSeek V4.1 自评得分远高于同为 V4.1 的子智能体
有人发现 DeepSeek V4.1 给自己打分挺高，给同样是自己扮演的子智能体打分却低不少，这个自评偏差挺有意思。
Fish Audio Drama 3 Preview Tops Open-Weights TTS Leaderboard
An open-weights model beating closed leaders like Eleven v4 on pronunciation while costing a fifth of the price signals real pressure on commercial TTS pricing. Builders producing audiobooks, dubbed content or multi-speaker agents can self-host the weights from Hugging Face or call the API, with 83-language support covering major Southeast Asian languages. The Elo gain of 101 points over Fish Audio's own S2.1 Pro shows the jump is internal as well as competitive.
Chinese contentDeepSeek V4.1 Flash 在 ARC-AGI-3 得分 3.3%，评估成本 $878
ARC Prize 实测了 DeepSeek V4.1 Flash，ARC-AGI-3 跑出 3.3%，成本不到 1000 美元，目前性价比最好的一份成绩单。
Oct 13
Chinese contentCodeVideo Model 项目发布：代码演示视频生成模型开源
一个叫 CodeVideo Model 的开源项目，做的是自动生成代码演示视频，项目页和教程都放出来了，做开发教程内容的可以看看。
Chinese contentV4.1 与 Sol 6.1 表现出色，Haiku 依旧不在小模型赛道
V4.1 和 Sol 6.1 这轮表现都很能打，尤其 Sol 6.1 出乎意料，还顺带聊了 Haiku 和 Fable 5.5 的定位。
Chinese contentRecraft Studio 演示 AI 头像全流程：V4.1 Pro 加 Nano Banana 加 Seedance 2.5
Recraft 演示了一条三步头像工作流，从 V4.1 Pro 出图到 Seedance 2.5 成视频，提示词全公开，直接照抄就能跑一遍。
Chinese content消息称 DeepSeek 将完成 1000 亿元融资，投前估值仍约 5000 亿元
DeepSeek 这轮融资从 500 亿直接翻倍到 1000 亿元，估值却没涨，新进来的全是吉利、中芯聚源这种产业资本，还能顺带看看它科创板上市的筹备细节。
FrontierCode 1.1 Leaderboard Launches on Artificial Analysis
The benchmark penalises behaviours conventional evals ignore, such as oversized diffs, out-of-scope file edits and unhelpful tests, so its ranking offers a different picture of coding ability than SWE-style test suites. Even the best model fails over 40% of tasks, meaning maintainer cleanup is still unavoidable. Teams choosing models by cost-performance should note GPT-6.1 Sol matching Claude Fable 5.1 at under a tenth of the per-task price.
Artificial Analysis launches leaderboard for Cognition's FrontierCode 1.1
The benchmark's merge-focused grading penalises the messy diffs and unhelpful tests that cheaper evaluations ignore, so leaderboard gaps here reflect real maintainer acceptance rather than raw test-passing. Leading models still fail over 40% of tasks, and nine tasks went unsolved by every model evaluated. Teams choosing coding agents can weigh GPT-6.1 Sol's 49% score at US$0.81 per task against Claude Opus 5.5's 52% at far higher cost, with open-weights options like MiMo-V2.6-Pro at 45% close behind.
Chinese contentAgent Arena 数据：开源模型 12 周追平 Claude Fable 5
LMArena 给出实测数据：Kimi K3 和 DeepSeek V4.1 Flash 用不到一成的成本追平了 Claude Fable 5，做 agent 选型前可以看看这份榜单。
Chinese contentMeta 发布 Muse Wonder 与 Muse Wonder Full，智能指数升至 52 和 56
Meta 六周内又发了新模型，智能指数涨到 56，代理类任务进步明显，比 Claude 便宜但比 GPT-6 贵，做 agent 的可以看看。
Oct 12
Chinese contentRecraft 推出 V4 Styles，可基于单张图像延伸新视觉方向
Recraft 上线了 V4 Styles，扔一张图进去就能延展出好几种新视觉方向，做设计探索挺方便，评论区还附了提示词。
Chinese contentOpenCode 实测：80 美元在 DeepSeek V4.1 Flash 上跑了 34670 次请求
有人实测 80 美元在 DeepSeek V4.1 Flash 上跑了 3 万多次请求，比 Command Code 还多，写代码跑量的可以看看这笔账。
Chinese content用 Codex + Deepseek-V4.1-Flash 以 JSON 生成可导出 SVG/PNG/GIF/MP4 的动画架构图
有人用 Codex 加 Deepseek-V4.1-Flash，让模型只写 JSON 就能自动生成会动的架构图，还能导出 GIF 和 MP4，画时序图、Sankey 图都行，做技术分享很实用。
Chinese content研究：SWE-bench 上通过测试的补丁近半数存在运行性能差异
四个前沿模型解 SWE-bench，通过测试的补丁一半有性能差距，模型性能排名还和解决率反着来，做工程的人可以看看。
Adaptive DHiSparse Speculative Decoding
This approach solves a critical bottleneck in long-context AI serving by enabling more efficient memory usage while maintaining high throughput. The technology allows models to handle significantly larger contexts without hitting memory limits, which directly impacts real-world applications requiring extensive context windows. Developers working with large language models should consider this optimization technique to improve performance in production environments.
Oct 11
Internalizer maps context to parameters for DeepSeek V4-Flash
The shared-expert placement means one hypernetwork can steer every layer's routed experts at once, which sidesteps the usual per-layer adapter sprawl and could make context injection cheaper at inference time. Teams running Mixture-of-Experts models such as DeepSeek V4-Flash or Qwen variants should watch whether the mapping survives fine-tuning and long-context loads before adopting it over LoRA.
Tsinghua, Oxford, Stanford Paper Finds LLMs Reason Even When Told Not To
The finding that a missing think tag or short reply does not mean a model skipped reasoning undermines a common evaluation shortcut. Teams benchmarking reasoning models or building answer-only constraints should re-check how compliance is measured, since suppressing visible reasoning costs roughly 15 accuracy points. The result spans 5 models, making it a broad behavioral pattern rather than a single-vendor quirk.
Open SWE routes tasks to cheaper models, cutting cost 64%
Cost-aware routing in the agent harness, rather than a single frontier model everywhere, is becoming a practical default for coding agents. Teams building multi-step agents can benchmark cheaper models per task type and keep premium models for escalation only. The DeepSeek V4.1 Flash versus Opus split shows a concrete tiering pattern worth copying.
