文档解析模型 OXR 把解析失败直接喂回训练循环,1.3B 参数就拿下 olmOCR-Bench 等三份基准第一,速度还是 GLM-OCR 的 3.05 倍。
Latest
DeepSeek 把跑 V4.1-Flash 的推理引擎 DeepInfer 开源了,基于 vLLM,想自建部署省算力成本的可以拿来直接用。
ElevenLabs 出了个语音转语音模型 v1 Max,0.57 秒就回话,Big Bench Audio 跑分 98.8%,做语音对话的可以看看。
Google 又憋了个 Gemini 4 Carbon,要在刚发的 Argon 上再升级,和 Anthropic、OpenAI 掰手腕,可以看看细节。
The fixed verifier design is the key detail: it keeps the agent's self-improvement honest by preventing it from altering how its own outputs are judged. Builders working on long-running agents can borrow the pattern of persistent history, versioned branches and independent review. The Épi-turbo comparison against Codex on FrontierOR offers a concrete cost-versus-accuracy reference point for teams automating solver generation.
DeepSeek V4.1 Flash 出了去审查版 GGUF,拒答率直接归零,MMLU 还从 86.0 涨到 86.6,想本地跑的可以去试试。
fal 公布了 FLUX 3 Max 的第三方 Elo 成绩,1,139 分进了图生视频第一梯队,做视频生成的可以看看这个榜单。
这期开源周报把 DeepSeek V4.1 Flash 1017、美国本土开源模型和中美监管动向一次讲清,想追开源进展直接看这篇就行。
Mistral 又出新东西了,这次是语音智能体整套方案,50 多种语言还能边聊边调工具,数据可以自己部署,搞语音方向的可以看看。
Mistral 出了个语音全家桶 Voxtral 3,转录、配音、语音对话都有,还能自己部署,做语音应用的朋友可以看看。
DeepSeek 把推理集群的账本公开了:日处理 155 万亿 token,GPU 时收入 38 美元,算下来毛利约 82%,这数字值得看看。
Legacy COBOL systems still run banking and government payroll worldwide, and a 97.9 percent migration accuracy claim suggests AI-assisted mainframe modernization is becoming practical rather than experimental. Teams maintaining COBOL or Fortran codebases should evaluate GPT-6.1 Sol against Anthropic's Claude and other coding assistants on their own migration test suites. The MultiLangBench scores also give enterprises a benchmark for comparing models beyond mainstream languages like Python.
Salesforce 出了个直接听语音的模型 Launa-Omni,不像一般语音 AI 先转文字,语调情绪都能留住。
Anthropic 把没发布的模型给 Citrini 用,Claude 自动跑了 300 个 sub-agents,从 31.8 万个候选里筛出 211 个,17 个真能结合毒素。
Voxtral 3 puts Mistral within striking distance of ElevenLabs and Grok on transcription accuracy while undercutting most premium TTS providers on price, making voice AI cheaper for production workloads. Builders choosing between streaming latency and accuracy can compare the 480 ms and 960 ms settings directly. Simba 3.2 at $6.6/1M characters remains the budget alternative for cost-sensitive deployments.
Krea 家新出的 Ultra 模式图像编辑冲到榜单第 5,比 OpenAI 和 Google 的模型都靠前,2K 编辑 0.27 美元一张,做图的朋友可以试试。
独立实验室 Krea 的 3 Ultra 图像编辑拿到 1,151 Elo,追平微软 MAI-Image-2.6,只差 Google 的 Nano Banana 2.1 一分,小团队也能和大厂掰手腕了。
The notable detail is the price-performance gap: Fibo 2.0 leads both leaderboards while costing less than half of Luma UNI 1 Max and Qwen Image 2.0 Pro, the closest commercial competitors. Teams building image generation or editing pipelines should test it via the Bria API or fal, but note the weights carry a non-commercial license requiring a separate Bria agreement for commercial deployment.
Krea 新模型图像编辑冲到榜单第 4,只差 Nano Banana 2.1 一分,还比 GPT Image 2 快 10 倍,单张最低 3 美分。
Krea 的新模型 Krea 3 把图像竞技场第一拿下了,Ultra 模式只有 OpenAI 比它强,Fast 模式还兼顾速度,做图像生成的可以试试。
DeepSeek 把自家跑 V4.1-Flash 的推理框架开源了,Apache 2.0 随便商用,还同时支持英伟达和昇腾,做部署的可以看看。
做法律 AI 或想选模型的看下,Legora 每月用真实律所案例测 17 个模型,还加了按任务类型筛选,比看总榜实用多了。
Bria 的 Fibo 2.0 登陆 fal 了,文生图加指令式修图,最多喂 4 张参考图,全是授权数据能商用,做电商图的朋友可以试试。
A 94.1% score above the human noise ceiling suggests behavioural-simulation systems like Engram 1.0 can outperform frontier LLMs at modelling subgroups, not just answer questions. Product and research teams building personalisation, market research or user modelling tools now have a public baseline to test against. The 601 private cases make leaderboard gaming harder, so results should hold up better than fully public benchmarks.
DeepSeek 新版 V4.1 Flash 1017 在多步网络安全任务上跑赢了开 Daybreak 的 GPT 6 Sol,测试帖细节可以看看。
国内团队 VAST 把一句话生成可玩 3D 世界的全流程开源了,规则和画面分开处理,想自己做生成式游戏的可以拿来跑跑看。
Claude 64小时自己搞出31.8万个蛋白质设计,实验室实测17个真的能用,AI科学家这回是真干活了。
The notable result is efficiency: matching frontier models on agent benchmarks while emitting far fewer tokens per task, which directly cuts inference cost for agentic workflows. Developers running local agent stacks can fit this open model on one RTX 4090 and get longer sessions via the MLA version. Teams choosing between open small models and GPT-6 Luna now have benchmark-level evidence that post-trained Qwen variants compete on most agent tasks.
Anthropic 新模型被内测者曝出能顺藤摸瓜找出 bug 背后的根因,不只改你报的那个错,写代码的可以关注下。
Tibo 网站倒计时显示 11 月 1 日 10 点,疑似 GPT-6.5 要来了,OpenAI 还没官宣,先当传闻看。
OpenBMB 把 35B 多模态模型塞进 iPhone,每秒 20+ token,比基线快 6 倍,代码开源可以直接上手玩。
Meta 开源了能塞进一张 RTX 5090 的 30B 智能体模型,工具调用和长程任务跑分压过同级 Qwen 和 Gemma,做本地 Agent 的可以试试。
The interesting signal is the trajectory comment: with heavy token budgets still unsaturated, frontier reasoning models may keep gaining from simply spending more compute per task. Practitioners tracking WeirdML results and reasoning-setting trade-offs can use this to decide which high-effort model configuration to benchmark next. The comparison between 1017 high and Grok 4.7 offers a concrete baseline for further testing.
DeepSeek 把自家用在生产环境的推理引擎开源了,还兼容华为昇腾 950 芯片,看推理优化的人可以拆开研究。
Cloudflare 出了个叫 Clairic 的模型,专门做文档数据提取,官方说比直接用 LLM 更快更省钱,处理文档多的可以看看。
Bindu Reddy 说 DeepSeek 性能赶上 Sonnet,价格却只有十分之一,缺点是速度慢但团队在修,性价比党可以关注。
Only the latest 2000 items are listed; for older ones see the daily archive