Chinese content四个 Gemini 4 Argon 智能体 27 小时自主通关《我的世界》
谷歌用四个 Gemini 4 智能体互相配合，27 小时自己通关了《我的世界》，还击败了末影龙，成本约 1620 美元，全程没人插手。
IT之家Source
Latest
Every published item, in order · 9305 items
Oct 19
Chinese contentThe Information：Anthropic 将于本月发布新模型
The Information 说 Anthropic 这个月要发新模型，亮点是能猜出你没说出口的意图、建议不那么套路，同期 Gemini 4 Carbon 和 OpenAI 的 Terra 也排队登场。
Nace AI launches Drex Max multimodal decision model
A decision model tuned for reading screens first could simplify computer-use and browser-use agents, where each step currently depends on interpreting the display before acting. Developers building such agents can evaluate it against Jev 1.14 on Mind2Web and the Decision Index. At US$0.06 per 1M input tokens, pricing is a concrete comparison point for teams balancing latency and cost.
MiniMax M3.1 Flash: Agentic AI at 1/3 the Cost
Flat-rate pricing at US$0.15 per 1M input tokens undercuts most frontier-tier APIs, making M3.1 Flash a serious option for cost-sensitive agent and coding workloads. One tester reports its front-end generation is the strongest among Chinese models. Builders in Southeast Asia can try it now via platform.minimax.io ahead of the open-weights release.
Chinese contentNace AI 发布 Drex Max 决策模型，支持图像视频输入
Nace AI 出了个专门做选择题的决策模型 Drex Max，能看截图给每个选项打分，比 Claude、GPT 这些模型便宜最多 167 倍，做浏览器智能体的可以试试。
Chinese contentAnthropic 新模型可举一反三：修一个 bug 时主动追问是否处理同类问题
Anthropic 新模型被曝能顺着 bug 找出整类问题，还会先问你修不修，来源是 The Information 的报道。
Chinese contentMiniMax 发布 M3.1 Flash：428B MoE 模型，输入 $0.15/百万 token
MiniMax 又上新了：M3.1 Flash，1M 上下文，输入才 $0.15/百万 token，跑智能体成本很低，10 月 23 日还会开源权重。
MiniMax-M3.1-Flash Debuts at #12 Among Open Models in Code Arena
A one-third cost cut plus a 3x speed gain while climbing 106 points on the WebDev leaderboard makes M3.1-Flash one of the most price-competitive coding models available via API right now. Developers building agentic coding or automation workflows can compare it directly against pricier frontier options at the flat US$0.15/US$0.60 rate, and open weights landing on 23 Oct will enable self-hosting.
Chinese contentMiniMax-M3.1-Flash 在 Code Arena WebDev 排名公开模型第12
MiniMax 的 M3.1-Flash 在 LMArena 网页开发评测里开源模型排第12、总榜32，想比较开源模型写代码水平的可以看看这个榜。
Tomorrow 08:00, the daily lands in your inboxTelegram channel or email. Same screening as this page, no ads.
Chinese contentMiniMax 发布 M3.1 Flash：价格为 M3 的三分之一，速度提升 3 倍
MiniMax 把 M3.1 Flash 做到 M3 三分之一价格、三倍速度，编程和视觉推理基准都好看，权重 10 月 23 日就开源，可以关注下性价比。
Google Research releases Harness² framework for agent self-improvement
The implication is that harness scaffolding, not model weights, is a lever for gains on unseen tasks, with a frozen model still improving when its surrounding code is edited. Developers building agent systems can apply the editor-composer-reflector pattern to their own pipelines. Code and the arXiv paper (2610.14224) are public, so teams can reproduce the benchmark setup directly.
OpenRouter reveals Space Bunny as MiniMax M3.1 Flash
A stealth release topping OpenRouter's leaderboard gives an unusually honest read on how a Chinese lab's model performs under real, unlabelled traffic. Developers can test MiniMax M3.1 Flash free for 72 hours and compare it directly with incumbent options on the same routing platform. The token volume also signals how quickly anonymous Chinese models can gain adoption in Western developer pipelines.
Chinese contentMiniMax M3.1 Flash 上线：推理速度提升近 3 倍，成本降低 70%
MiniMax 新发的 M3.1 Flash 又快又便宜，写代码和操作电脑的基准分数都很能打，API 价格每百万 Token 只要 1 元起。
IT之家Source
MiniMax M3.1 Flash Gets Day-0 vLLM Support
Day-0 vLLM support means teams can self-host the new MiniMax model immediately without waiting for community adapters, and NVFP4 quantisation cuts serving cost for inference-heavy agent workloads. Developers running coding or automation agents on their own infrastructure should evaluate it against comparable open serving stacks. MiniMax is a China-based lab, relevant to readers tracking Chinese open models.
Chinese contentMiniMax M3.1 Flash 公布官方基准，价格对标 DeepSeek V4.1 Flash
MiniMax 放出 M3.1 Flash 官方跑分，6 项基准里 4 项输给 DeepSeek V4.1 Flash，但高峰期价格便宜一半，选谁看场景。
Chinese contentMiniMax 发布 M3.1 Flash：价格为 M3 的 1/3、速度 3 倍
MiniMax 出了 M3.1 Flash，价格是 M3 的三分之一，速度快 3 倍，跑分接近旗舰水平，10 月 23 日还会开源权重。
Chinese contentMiniMax 发布 M3.1 Flash：价格降至 M3 的 1/3
MiniMax 新发的 M3.1 Flash 又便宜又快，编码和视觉推理分数都挺高，比 M3 省三分之二的成本，可以试试。
DeepSeek-V4.1-Flash-1017 builds full projects from scratch
Self-feedback reinforcement learning that targets whole-project output, not single-file answers, signals a shift in coding benchmarks toward multi-file, runnable deliverables. Developers evaluating open-weight models for end-to-end scaffolding can compare V4.1-Flash-1017 against Claude and GPT coding models on game and 3D tasks before committing to a stack.
Chinese content网友推测 DeepSeek 新模型 DSV4.1F-1017 在 Terminal-Bench 4.0 得分 42.3
有人发现 DeepSeek 这个代号奇怪的模型在 Terminal-Bench 4.0 上拿了 42.3，还比 GLM-5.3 强，等 Artificial Analysis 复测。
Chinese contentMuse Wonder Full 与 Gemini 4 Argon 在 CancerBench 上均得零分
两款新模型上线 CancerBench 评测，结果都拿零分，还并列第一和倒数第一，挺有意思的一条测试动态。
Chinese contentArena 第 42 周盲测榜：kimi-k3-max 稳居代码榜第十
这周盲测榜出来了，kimi-k3-max 代码榜冲进前十，小米 mimo-v2.6-pro 也往前赶，想选模型可以对着榜单比比价格再下手。
IT之家Source
Chinese content谷歌让AI Agent进入真实生物实验室，闭环发现癌症靶点结合分子
谷歌把Agent从电脑里设计分子推进到真实实验室，机器人做实验、AI自己改方案，还真的找到了癌症靶点的结合分子。
Chinese contentACE Robotics 发布 9B 具身模型 ACE-Brain-0.6，主打空间视觉定位
ACE Robotics 拿 100 万条空间数据微调了个 9B 模型，自称八项测试里六项赢 GPT-6 Astra，不过权重还没开源。
Chinese content大晓机器人发布具身基模型 ACE-Brain-0.6，六项空间感知评测超 GPT-6 Astra
商汤系公司大晓机器人发了具身基模型 ACE-Brain-0.6，基于 Qwen3.5-9B 微调，六项 Grounding 评测分数超过了 OpenAI GPT-6 Astra，做机器人方向的可以看看评测细节。
IT之家Source
Chinese content用户用 Claude Opus 5.5 纯代码生成古希腊世界，不使用任何素材
有人让 Claude Opus 5.5 造了个古希腊世界，全程没用素材，全是代码画出来的，看看效果如何
Chinese contentCoT 推理链越来越长，OpenAI 的 token 消耗量被认为居首
讨论各家推理模型 CoT 越写越长的问题，作者点名 OpenAI 的 token 消耗比 DeepSeek 还夸张，关注 API 成本的可以看看。
Chinese contentDeepSeek 新模型在 GDPval 基准上超越 OpenAI
DeepSeek 新模型在 OpenAI 自家的 GDPval 基准上把开源和中国模型都拿了第一，还比同分模型快，数据点很硬。
Chinese contentDeepSeek V4.1 Flash 1017 上线 Together AI，聚焦完整项目构建
DeepSeek 新的 V4.1 Flash 1017 在 Together AI 能直接用了，专攻从零写完整 Web 应用和游戏，价格没涨，写代码的可以试试。
Chinese contentDeepSeek V4.1-Flash-1017 被评幻觉最少但知识面倒退
DeepSeek 新版 V4.1-Flash-1017 的第三方点评来了：幻觉控制最好，但知识和 LCR 倒退，还被 Sol 压着打。
Chinese contentDrugHarness 发布：可进化的药物发现智能体框架
这个智能体框架专门给药物研发用，会自己从实验轨迹里学新技能，拿 Qwen 跑出来的准确率能提近 19 个百分点，还能和真实湿实验对上。
Chinese content网友用 Claude 创作整首歌《Don't Let Me Scroll》，仅保留人工写的歌词
有人只写了歌词，剩下的曲子和演唱全丢给 Claude 完成，成片就贴在推文里，点开就能听，挺有意思的实验。
Chinese contentMiniCPM-V 4.7 登上 Hugging Face 热榜第二
OpenBMB 的 MiniCPM-V 4.7 冲上 Hugging Face 热榜第二，35B-A3B MoE 手机端 3GB 内存就能跑，想玩端侧多模态的可以试试。
Chinese contentDeepSeek 周六发布 Flash 档模型，评测机构结果待出
DeepSeek 周六悄悄上了个 Flash 档模型，先别急着下结论，等 Artificial Analysis 和 ARC 这些评测跑完再说。
Alibaba Tongyi Open-Sources OpenGCUA Computer-Use Agent Stack
The interesting shift is that code execution as the agent substrate lets weaker models close the gap with frontier ones by writing programs instead of clicking precisely, which changes how agent-vs-model capability should be evaluated. Developers building computer-use agents can reuse the harness, SDK and macOS app directly and compare against GUI-only and hybrid designs that the paper treats as special cases. The OSWorld-v2 results give a concrete head-to-head of GPT-6 Astra, Claude Opus 5 and Chinese models like GLM and Kimi under the same harness.
D3 decision models top Jev Decision Index with vLLM recipes
Decision models that return scores and classifications instead of long generations can cut serving costs for routing, moderation and scoring workloads. Teams running their own inference stack can start with a vLLM nightly command and the merged recipes to benchmark D3 on their own hardware, including AMD MI325X GPUs. Self-hosting keeps data residency and latency budgets under direct control.
Mistral open-sources Timestral for multivariate time-series forecasting
A 5M-parameter model that tops zero-shot forecasting on GIFT-Eval means teams can run probabilistic forecasting on CPUs without GPU infrastructure, cutting deployment costs for demand or capacity planning. Engineers building forecasting pipelines should compare Timestral against TimesFM and Chronos on their own data. The 21-quantile output makes it practical for risk-aware decisions, not just point forecasts.
Only the latest 2000 items are listed; for older ones see the daily archive