Google 悄悄更新了图像模型 Nano Banana 2.1,还没基准数据,帖子里也在问它能不能打过 GPT Images
Latest
有人拿 8 个决策模型打 Tetris,Perplexity Decider 稳居第一,还能自己在 playground 里复现,挺好玩的。
Mistral 这次真追上来了,排名压过所有美国开源模型,只差 DeepSeek 和 Kimi 一点,欧洲开源有看头。
Stanford 的 Marin 开源模型训到一半了,535B 总参数、每次激活 23B,Percy Liang 还分享了训练中途的踩坑心得。
Cloudflare 的 Clef Flash 上 OpenRouter 了,点进 Playground 就能直接试,不用自己搭环境。
A 740M embedding model that beats larger specialist models makes high-quality multimodal search feasible on consumer hardware without cloud calls. Developers building local RAG pipelines can download the Apache 2.0 weights from Hugging Face or Kaggle today and combine them with Gemma 4 for a fully on-device stack. Compared with larger commercial embedding APIs, it trades scale for privacy and zero inference cost.
MIT CSAIL 搞了个叫 DREAM 的模型,一个模型又能看图又能画图,还会自己挑草稿,生成还快了约 10%。
有人在 Copilot 365 Premium 里刷出了 GPT-6.1 Sol,比官方公告还早,想第一时间尝鲜的可以去看看自己的账号有没有。
DeepSeek 的 V4.1 Flash 跑分出来了,ARC-AGI-2 拿下 72.9%,比上一代高了 11.5 分,不过每任务成本贵了两倍多,看你怎么权衡。
Mistral 放出 1T 总参、49B 激活的 Le Chonk,Simon Willison 亲测能画出鹈鹕 SVG 了,10 月底开放权重。
Mistral 的新旗舰进了 Arena,1T 参数只激活 49B,开放权重 10 月底就能下载,想去投票试试吗。
Real-time execution, not turn-based pausing, is the hard part — the game keeps running while each agent plans, testing latency-aware decision-making beyond scripted bots. Builders working on agents can study the four-tool-per-agent design and fog-of-war state sharing as a blueprint for competitive multi-agent environments. Compared with classic StarCraft AI research like AlphaStar, this setup is a lightweight self-assembled stack rather than a lab-scale training effort.
复旦、腾讯混元和浙大联合开源了 Prism,能生成 2K 分辨率且声音和画面同步对齐的视频,训练速度还快了 2.5 倍,做视频生成的可以去看看。
DeepMind 前员工做的单一模型,画图改视频控制机械臂全包,延迟砍半,不用再等十来秒了。
Mistral 放了个 1T 参数的大家伙,编码分数把 Reflection 的 Beam 甩开 17 个点,月底开源权重,可以等等看。
Mistral 放出 1T 参数的多模态模型,激活只有 49B,API 现在就能用,10 月底开源权重,欧洲目前最能打的一个。
The cheaper, faster Turbo variant beating its own flagship in blind human preference voting signals that latency and cost, not raw quality, now decide TTS purchases. Teams building voice agents or customer service bots can pick Turbo for roughly half the price of Eleven v4 while still topping the arena. Its US$40/1M characters undercuts Cartesia's Sonic 3.6 at US$49, though Gemini 3.8 Flash TTS remains far cheaper at US$16.49.
Mistral 放出 1T 总参、激活只有 49B 的 Mistral Large 4,视觉定位据说比 GPT 和 Claude 还强,月底还有开源权重版可以自己部署。
A cryptic post from HuggingFace's CEO pointing at Mistral's model hub page is a classic pre-release signal, and the French phrasing fuels speculation about a new open-weight model. Practitioners tracking open-source models should watch the mistralai page on HuggingFace for the drop. Mistral's releases typically land as open weights, making them direct alternatives to Llama and Qwen for self-hosting.
Mistral 的新一代 Large 4.0 已经能在 OpenRouter 上调用了,做模型对比或接 API 的可以直接试试。
Mistral 这次真支棱起来了,1T 参数只激活 49B,用 4000 块芯片就训出能跟 GLM-5.3 掰手腕的模型,欧洲总算有能打的了。
Mistral 拿出 1T 参数的 Le Chonk,激活只有 49B,跑分号称欧洲开源第一,还能在视觉定位上压过闭源模型,10 月底开源权重,可以关注。
Mistral放出Mistral Large 4,1T参数只激活49B,跑分被DeepSeek V4.1 Flash压了一头,10月就开源,可以先试试API。
有人让 Claude 连续干 18 小时,独立做出了一个能跑的 Windows XP 复刻版,视频里的完成度相当惊人。
Mistral 放出 1T 参数的 Le Chonk,激活才 49B,视觉定位还能打赢闭源模型,10 月底开源权重,开源党可以蹲一下。
欧洲的 Mistral 出了个 1T 总参数但只激活 49B 的多模态模型,视觉理解敢跟闭源模型叫板,权重 10 月底开源,可以蹲一下。
Mistral 放出了 1T 总参、49B 激活的多模态模型预览版,说是欧美最强开源权重模型,月底连权重一起放出来,做开源部署的可以盯着。
Mistral Large 4 在多个专业领域基准测试中表现优异,金融和法律任务得分亮眼。
有人在 DGX Spark 上用自写的 vLLM 补丁跑 DiffusionGemma 和 Jev 对比,vLLM 官方都转发了,想看扩散模型推理性能的可以点进去。
The hallucination rate fell from 44% to 38% while accuracy rose from 18% to 29%, meaning the model genuinely knows more rather than abstaining more often. At US$0.99 per task it sits between DeepSeek V4.1 Flash (Max, US$0.32) and Gemini 3.8 Flash (High, US$1.24), and 16% fewer output tokens partially offsets the 4x price increase. Teams building agentic workflows who need open weights at a 1M context window should compare it directly against GLM-5.3-Flash.
Vals AI 派了 90 多个 Claude Opus 5.5 智能体做材料模拟,3 天找到两个磁性半导体候选,其中一个还是 1999 年就合成过的老材料。
A video model shipping with audio and lip-sync baked in, under a permissive MIT license, is rare — most rivals keep weights gated or restrict commercial use. Builders can immediately self-host it via vLLM-Omni without waiting for community ports. Compared with closed alternatives like Sora, this gives full control over inference stack and cost.
Bindu Reddy 亲自测评后说 DeepSeek 跑个人智能体又好又便宜,单用户推理成本几乎为零,想省钱搭智能体的可以看看。
Only the latest 2000 items are listed; for older ones see the daily archive