腾讯 Hunyuan 的 Hy4 在 OpenDesign 设计基准排第 4,只差 GPT-6.1 Sol 0.1 分,成本还低 41%,做设计的可以看下榜单。
Latest
一个 594B 但只激活 17A 的模型,68 小时就训完 12T tokens,还 beat 了 Muse Spark,训练拆解值得细看。
OpenAI 新出的 GPT-6 Cyber 专攻找漏洞,成绩比 GPT-5.6 Cyber 高了一截,还帮 Mozilla 挖出 81 个 Firefox 漏洞,安全圈可以关注。
The gap between Hy4 and the next open-weights model is nearly 8 points on Pass @1, which resets assumptions about where free-weights agent performance stands relative to closed models. Teams selecting models for professional-work agents—consulting, finance, legal workflows—can use the Mercor APEX-Agents leaderboard to shortlist candidates by domain rather than headline scores. Comparisons against Kimi K3.1 Pro and Qwen3.8-Max make it a practical benchmark for China open-weights options.
腾讯 Hy4 用的人太多,WorkBuddy 直接扩了十倍,Agent 能力还拿下开源第一,写 Agent 的可以关注下。
webAI 放出了一个 8B 的小模型 TwIL-8B,专攻数学和编程,跑分居然赢过了大几倍的 Gemma 4 31B,本地部署党可以看看 HuggingFace。
Fish Audio 的转写模型在 Speko 真实对话测试里词错率 11.5% 排第一,Qwen3-ASR、ElevenLabs、OpenAI 都被比下去了。
腾讯的 Hy4 在 Agent Arena 智能体榜单拿了中国模型第一,12000 多场真实任务测出来的,单任务成本才 0.57 美元。
淘宝 MNN 团队开源了一个能在手机上完全离线跑的 0.3B 模型,iPhone 18 Pro 实测 350 token/s,比 Qwen3-0.6B 更小更快,做端侧应用的同学可以直接拿去部署。
外科手术版的 DINOv3,用近 2900 小时手术视频训练出的开源视觉编码器,20 个手术任务全胜,做医疗影像的可以拿来当 backbone。
Google 新出的 Gemini Omni 1.2 Flash 有 9 个演示,视频能改角色换场景还能保留表演,跟之前一次性生成很不一样。
The architectural split between 3D state prediction and rendering avoids the drift that plagues video-frame-recurrent world models, and the laptop-scale latency shows the approach runs on consumer hardware rather than data-centre GPUs. Developers building playable-world or interactive-simulation demos can download the code and weights and extend the framework to other games or rule-based gameplay. It compares with video-only world models like Genie-style systems by trading appearance realism for spatial persistence and control.
The upgrade comes from post-training and perception rather than scale, which means the same 770B base now handles screenshots, charts and full desktop workflows — a shift that matters more for agent builders than raw leaderboard chasers. Missing third-party evaluations mean teams should run their own tests before switching, and the free quota until 15 Nov 2026 makes a hands-on comparison against DeepSeek or Qwen low-cost to attempt.
一个 65B 的小模型在 TerminalBench 上打赢 816B 的大块头,还支持 1M 上下文,写代码的可以试试权重。
百度把 ERNIE-5.5-Flash-Base 的预训练检查点直接开源了,280B 参数只激活 14B,想自己微调基座模型的人可以去看看。
腾讯混元 Hy4 冲到 Agent Arena 中文第一,单任务成本比 Kimi K3.1 Pro 便宜六成多,分数还几乎持平,跑智能体任务可以试试它。
Thinking Machines 的新模型 Inkling 2 Small 在 WebDev 基准拿了 1654 分,比上一代高 245 分,权重开源后将是美国开源第一,还支持微调。
OpenAI 出了专门做网络安全的模型 GPT-6 Cyber,攻防测试成绩超过自家 GPT-5.6 Cyber 和 GPT-6 Astra,但官方也标了高风险。
蚂蚁发了 Ling-3.2-flash,560B 参数原生多模态,自测三项基准跑赢 Claude Opus 5(max),API 前两周免费,可以试试。
Particle-1 下一款模型在路上了,有开发者实测代码审查场景比 Deepseek v4 Flash 快 8-10 倍,可以看看这个说法。
谷歌的视频模型又升级了,能一次生成 15 秒还降价三成,改镜头不用重做,做广告片的可以试试。
Google DeepMind 的新视频模型,片段从 10 秒加到 15 秒,口型和角色跨镜头一致性都改进了,做短视频的可以试试。
Stanford AI Lab 校友Eric Zelikman说他们本来想训练模型更好地跟人协作,结果顺手做成了很强的多智能体编程模型,方向挺意外的。
百度把 280B 的 ERNIE 5.5 Flash 基座模型在还没做后训练前就放出来了,想自己微调、做后训练研究的开发者可以拿来玩。
7 个人做出来的图像模型冲到 Arena 编辑榜第 3,Krea CEO 亲口讲训练过程,还有开源计划。
Voice Arena 做了个按语言和场景拆分的语音识别基准,第一个数据集是 5 种语言的真实双人对话,做语音智能体的可以去看看排行。
METR 拿到 Anthropic 内部未发布模型的数据,算出 1.9 倍加速率,比 Haiku 5.5 报的 1.7 倍还高,看看这些闭门模型到底多强。
OpenRouter 的神秘模型 Crouton Alpha 修好了提前罢工的 bug,免费还能塞 1M 上下文,跑智能体任务可以再试试。
Google 推出了 Gemini Omni 1.2 Flash,主打跑得快、价格低、视频支持更长,如果你的 Agent 需要处理视频可以留意。
METR 透露 Anthropic 上月训完的新模型,全部测试基准首次都过 50%,是这批模型里分最高的。
Anthropic 新模型在 ECI 上跑出 178.6,比原来第一的 Fable 5.1 高出 12 分,顶部模型平时就差几分,这个拉开幅度不小。
Google 把 Gemini Omni 1.2 Flash 的图生视频跑分放上 Arena 了,直接冲到第一,做视频生成的可以看看排名再选模型。
Google 的视频模型又更新了,登顶文生视频榜,能生成 15 秒的片子,API 还降价 30%,做视频的可以试试。
Voice Arena 搞了个语音识别基准,用真实对话测 ASR,不是念稿子,首套覆盖 5 种语言,做语音的可以去刷榜了。
A 4B model beating larger competitors on table and chart parsing changes the cost math for large-scale document pipelines, since inference spend scales with parameter count. Teams building RAG or data-extraction systems over PDFs, scans, and reports can try it immediately through LlamaParse's cost-effective and agentic modes. Compared with Mistral OCR 4.1, it targets the same workload at a claimed lower price point.
An 8.8B model matching near-frontier coding and math quality at consumer-hardware size changes the calculus for local inference, where running Claude- or GPT-class ability previously required cloud APIs. Developers building offline or cost-sensitive tools can deploy it from a single quantized 4.98 GiB file on commodity machines. Its efficiency-trained reasoning also cuts token counts, which matters for anyone benchmarking small models against Qwen3.5-9B or Gemma 4 31B.
The design borrows Lean's philosophy for AI math — verifiable, tamper-proof grading — and applies it to open-ended science, making contamination and benchmark gaming far harder. Researchers comparing model claims on scientific reasoning should treat SLE's 21.28% figure for DeepSeek V4.1 Flash as a more contamination-resistant signal than conventional Q&A benchmarks. The low score also sets a sober baseline for how far agents remain from genuine discovery.
Robustness to scene changes, not raw benchmark scores, is what separates deployable robot models from lab demos, and V0.5 posts a much smaller drop than π0.5 under distribution shift. Robotics teams evaluating open-weight manipulation models can benchmark V0.5 against π0.5 in their own cluttered environments. Its base on the Wan2.2 video generator also shows video world models doubling as action predictors.
Only the latest 2000 items are listed; for older ones see the daily archive