Chinese contentSGLang 联合 NVIDIA、AMD 为 MiniMax M3.1 Flash 提供推理与 RL 支持
MiniMax M3.1 Flash 发布当天就能用 SGLang 跑推理和 RL 训练，NVIDIA、AMD 卡都适配，自己部署的话可以关注下。
Latest
Every published item, in order · 9305 items
Oct 23
Chinese contentMiniMax-M3.1-Flash 获 NVIDIA NeMo 首日微调支持
MiniMax 的 M3.1-Flash 发布当天就能用 NVIDIA NeMo 做微调了，要做 post-train 的开发者可以省去等适配的时间。
Chinese contentMiniMax 开源 M3.1 Flash，总参数 428B、激活 23B
MiniMax 把 M3.1 Flash 的权重放出来了，428B 总参数激活 23B，还做了稀疏注意力和 W4A4 量化，上线三天就跑了超 100 万亿 Token，本地部署和私有化场景可以看看。
IT之家Source
Chinese contentMiniMax M3.1 Flash 开源权重，428B 参数每次激活 23B
MiniMax 把 M3.1 Flash 权重放出来了，428B 只激活 23B，代理和编码基准分都贴出来了，想本地部署的可以去 Hugging Face 拉 weights。
Chinese contentV4.1 Flash-1017 在 ARC-AGI-2 上略超 Gemini 3.1 Pro Preview
V4.1 Flash-1017 在 ARC-AGI-2 上小胜 Gemini 3.1 Pro Preview，想看推理模型基准对比的可以看看这组数据。
Artificial Analysis Coding Agent Index v2.0: Claude Opus 5.5 Leads
The v2.0 changes make rankings harder to game: private test sets, anti-solution scanning and multi-vote LLM judging push mean scores down 1.7 points across 21 shared configurations. Teams choosing a coding agent should note that Opus 5.5's edge is concentrated on FrontierCode (56.0% versus 52.6% for DeepSeek V4.1 Flash), while 15 of 24 configurations cluster between 43% and 49%, so harness choice can matter as much as model choice.
AA-PersonalityBench links model slop to intelligence gaps
Slop output is only loosely tied to raw capability, so model selection for writing or content pipelines should weigh the Slop ranking separately from the Intelligence Index. Teams building customer-facing copy or summarisation tools can use these two scores as a joint filter instead of assuming a smarter model reads better. Qwen3.8 Max ranking high on both lists is a useful data point when weighing it against Gemini 4 Argon or GPT-6 variants.
Chinese content开源模型 MemT 1.0 发布，能跨对话复用记忆
一个叫 MemT 1.0 的开源模型，能记住你昨天的反馈并用在今天类似任务上，靠的是 MoS 多存储专家架构，和普通 MoE 不一样。
Chinese contentFLUX 3 Max：12.45 秒生成 10 秒视频，速度超 Seedance 2.5 二十倍
fal 把 FLUX 3 Max 调校到 12 秒出 10 秒视频，比 Seedance 2.5 快 20 多倍还同分，做视频生成的可以看看他们怎么优化的。
Chinese contentCanva 发布自研图像模型，DesignArena 排名仅次于 OpenAI
做设计工具的 Canva 也自己训模型了，图像生成榜排名第二，先把排版和文字规划好再渲染，改图只动改的那块。
Tomorrow 08:00, the daily lands in your inboxTelegram channel or email. Same screening as this page, no ads.
AWS Priming Converts Transformers Into Efficient Hybrid Models
Priming makes hybrid architectures practical by skipping full pretraining, which could shift how teams convert existing checkpoints into cheaper-to-serve models. Engineers deploying 8B-32B scale models can use the released weights to cut inference memory and cost immediately. It competes with from-scratch hybrid efforts like those from other labs at a fraction of the compute spend.
Chinese contentMicrosoft 称其 AI 诊断工具准确率可达人类医生 4 倍
Microsoft 拿出诊断 AI，在 304 个真实医学病例上准确率约 80%，人类医生只有约 20%，差距大到离谱，来看看它怎么做到的。
CritPt v2 Benchmark Launches with Agentic Code Execution
The 54-point gap between GPT-5 and GPT-6 Astra in about a year shows how quickly frontier models are absorbing research-level physics when given code execution. Tracking these scores before the late October Intelligence Index update helps researchers and benchmark watchers pick models for scientific workflows. Compare against Claude Opus 5.5 at 78% when choosing a frontier model for agentic computation tasks.
Chinese contentDeepSeek V4.1 Flash 1017 在 ARC-AGI-2 上得分 78.1%
DeepSeek 又刷了 ARC-AGI-2，V4.1 Flash 1017 拿到 78.1%，每任务才 $0.16，分数和成本都很能打。
Chinese contentIdeogram 发布 Vector 1.0：生成可分层编辑的 SVG 矢量图
Ideogram 出了个 Vector 1.0，能直接生成 SVG 矢量图，每个元素还是独立图层，能随便拖动旋转，做设计的可以试试。
Chinese contentIdeogram 发布可直接生成矢量图片的模型
Ideogram 出了个能直接生成 SVG 的模型，放大不失真，还能让 Agent 用代码改图，做网页和海报的可以试试，同类还有 Recraft 和 Arrow。
Oct 22
Chinese contentRecursive 发布首个模型 Flint：40B MoE 编码模型由 AI 自主研发
前 DeepMind 团队做的 40B MoE 编码模型，整个研发流程由他们的 Eureka Machine 自主跑完，只花了 4 个月。
DeepSeek V4 Pro 0813 ties for second on CVE rediscovery benchmark
Vulnerability rediscovery is becoming a measurable benchmark for frontier models, and a Chinese model matching a top US lab result suggests the capability gap on security-relevant reasoning has narrowed. Security researchers and red teams can weigh these pass@3 scores when picking models for bug-hunting workflows. The poster argues the result reflects data scale rather than copying a competitor's outputs.
Chinese contentAA-PersonalityBench：Anthropic 模型在信息密度与文风指标上领先
Artificial Analysis 测了各家模型的性格指标，Claude 全系在信息密度和文风两项都排最前，写文案的可以参考这个榜。
Chinese content蚂蚁数科发布决策模型 LingDT-Decision-1.0
蚂蚁数科发了决策模型 LingDT-Decision-1.0，号称准确率超过微软和 Jev 的模型，做决策类任务可以关注下。
Chinese contentLing 发布 3.1-Flash，被类比为 Kimi K3-Flash
Ling（蚂蚁系）发了 3.1-Flash，有人拿它和 DeepSeek V4.1-Flash、Kimi K3-Flash 类比，可以看看这条讨论。
Chinese content蚂蚁数科发布决策模型 LingDT-Decision-1.0，Token 消耗降低超 90%
蚂蚁数科出了个专门做判断的决策模型，不做文本生成，Token 能省九成、快 10 倍，做客服风控的可以试试。
IT之家Source
Chinese contentRecursiveSI开源flint模型：40B总参数、2.5B激活，面向编程任务
RecursiveSI放出了开源编程模型flint，2.5B激活参数跑出比Qwen3.6 35BA3B更高的分数，还带262K上下文，写代码可以试试。
vLLM Ships Day-0 Support for Sage-A1.5
Day-0 support means the model can be self-hosted on existing vLLM infrastructure immediately, without waiting for community integrations. With only 3B active parameters, it targets teams running agentic workloads on modest local hardware who want an alternative to larger dense models. Developers comparing open-source MoE options for on-premise serving should evaluate its throughput against Qwen and DeepSeek MoE checkpoints.
Chinese content米哈游联合清华、北大等高校开源 GameScaling 游戏模型，覆盖 0.8B 至 27B
米哈游和清华北大把游戏操作模型开源了，能自己玩《我的世界》和《毁灭战士》，最小 0.8B 单卡就能跑，可以拿来折腾。
IT之家Source
Tencent Hunyuan releases WorkForge agent training framework
Training agents on verifiable, evidence-grounded tasks addresses the core bottleneck of RL for long-horizon work: reward reliability. The 16.7K-environment scale makes this one of the larger open efforts in agent environment construction. Engineers building work or office agents can study the dual-verifier design for their own training pipelines.
Chinese content中电信星辰实验室开源意图理解模型 Xing4.0-29B-Intent
中电信星辰实验室开源了 29B 的意图理解专用模型 Xing4.0-29B-Intent，做对话系统和意图识别的可以拿来试试。
Muse Glimmer 30B: Meta's distilled agent model runs on one consumer GPU
Distilling from a large teacher into a small model specialised for long tool-use chains may matter more than raw size for local agents, a trade-off the OSWorld gap makes visible. Builders running agents on local hardware can use it as a drop-in for multi-tool MCP workflows without cloud dependency. Compared with Qwen3.6-27B, it wins on tool orchestration but loses on general OS-level tasks.
Chinese contentOpenAI Patch the Planet 用 GPT-5.6 Sol 发现 PHP 核心 PHP2Shell 漏洞
OpenAI 用 GPT-5.6 Sol 挖出了影响全球 70% 网站的 PHP 核心漏洞，还和另一团队协同修了两个月，AI 安全研究的新玩法。
Chinese contentIBM 发布 Granite Speech 5.0 2B：开源语音识别与翻译模型
IBM 开源了个 2B 的语音模型 Granite Speech 5.0，Open ASR 榜单第二，英文 WER 只有 5.68%，商用免费拿去部署转录都行。
Chinese content科大讯飞发布语音基座大模型 SparkAudio 1.0，已接入讯飞输入法 16.0
讯飞把语音识别、翻译、情绪理解打包成了一个基座模型，输入法 16.0 就能中英混说，配套的录音环 2 小时录音 2 分钟转完。
IT之家Source
Chinese content讯飞星火 X2.6 发布：全国产算力训练，293B-A30B MoE 架构
星火 X2.6 不靠 N 卡训练，全用国产芯片，还把上下文做到 1M，输入定价 0.9 元/百万 tokens，端侧 4B 模型免费开源，可以试试。
IT之家Source
Chinese contentUnderdog Bark 3.0：仅 163 MB 的语音转文字模型
163 MB 的语音转写模型，M 系列芯片上半秒转完 5 分钟音频，错字率还比 Whisper 低，Clean 版连“嗯啊”都帮你删了。
Chinese content韩国魁匠团开源 Raon-1-mini-preview：29B 参数 MoE 模型
魁匠团这个做吃鸡的公司居然也开源模型了，29B 参数 MoE，韩英双语能力不错，感兴趣可以去试试。
Chinese content英伟达开源检索系列模型 Nemotron 3.5 Parse、Embed 与 Rerank
英伟达把检索三件套 Parse、Embed、Rerank 全开源了，扫描 PDF 和图表都能解析，准确率上去了成本还降了 14%，做 RAG 的可以试试。
Chinese contentNVIDIA Kumo Relational 在关系数据上超越监督学习，无需训练
NVIDIA 出了个 Kumo Relational 模型，关系数据不用训练直接预测，还比监督学习效果好，做数据分析的可以看看。
Only the latest 2000 items are listed; for older ones see the daily archive