与「MODEL」相关的搜索结果

The Evolution of SGLang, 2024 → ∞: from a caching trick to the inference engine behind 330+ organizations | 2026 Under three years, twenty milestones: from RadixAttention (Jan 2024) to SGLang v1.0 and the SGLang Consortium (Nov 2026). Today SGLang is used by 330+ organizations, supported 217 models on day 0, and runs on hundreds of thousands of GPUs. Music only. Every frame and every note generated in code. Sources: LMSYS blog and @sgl_project (Nov 12, 2026), @lmsysorg (Nov 11, 2026), @radixark (Nov 12, 2026), SGLang Summit 2026 timeline wall and keynote slides. AI-generated content (视频由 AI 生成).

joder, GitHub está lleno de cosas gratis que casi nadie está usando. los 10 repos que están explotando este mes, de lo más fácil a lo más técnico: lo que puede usar cualquiera 1. VoiceStudio: clona voces, dobla vídeos y transcribe en 646 idiomas, gratis y sin salir de tu ordenador (57K ⭐) 2. ppt-master: le pasas un PDF o un Word y te devuelve un PowerPoint de verdad, editable y con animaciones (59K ⭐) 3. gods-eye-view de @bilawalsidhu : un simulador de satélite espía en tu navegador, con aviones, barcos y cámaras reales sobre un globo 3D (50K ⭐) 4. REA: le enseñas a tu IA una función de otra app y te explica cómo está hecha, sin tener el código (65K ⭐) 5. diagram-design de @cathrynlavery : pides un diagrama con palabras y te lo dibuja con tu marca, 44 tipos (49K ⭐) 6. paperclip: como tener tu propia empresa de agentes autónomos trabajando para ti, la app open source para dirigirlos como si fueran tu plantilla (99K ⭐) para que Claude Code trabaje mejor 7. skills de @mattpocockuk : las que usa a diario un ingeniero de verdad, pequeñas y fáciles de adaptar (284K ⭐) 8. ECC: skills, memoria y seguridad en un pack para Claude Code, Codex, Cursor y OpenCode (276K ⭐) 9. andrej-karpathy-skills: un solo archivo con 4 reglas sacadas de lo que Karpathy dice que hacen mal los modelos programando (218K ⭐) 10. claude-mem: guarda lo que hizo tu agente, lo resume y se lo pasa a la siguiente sesión, para que no empiece de cero (99K ⭐) compártelo si te ha parecido interesante ♻️ y guárdalo antes de perderlo 🔖

Built at NEAR HACKATHON: X3 Force Field — a typed .x3 language + compiler for AI-agent spending rules. A NEAR Guard checks signed terms, not model promises. 25 USDC moved Solana→NEAR on mainnet under owner-approved limits. @NEARProtocol #NEARHackathon#

We ran Cybergym on every Kimi K3.1 Pro available online so you don't have to. Compared to unrestricted DeepSeek v4.1 1017 Flash, unrestricted Kimi K3.1 Pro, fine-tuned for offensive cyber, outperformed it with a 69.7% score on CyberGYM with the same harness. Combine this with a proper harness like PentAGI, ARTEX or similar. 1. Open weights available: It provides sovereign cyber AI access for companies, available for air-gapped deployments, trial possible on and on 2. This is almost same capability with Claude Mythos 5.2's safety guardrails removed version available for $999 per month for unlimited tokens. 3. This model is advertised as cyber data removed from it but actually it's we fine tuned it with our proprietary data surpassed the best second available open weights model built for cybersecurity on huggingface by ~10% 4. Our server health is public you can check here:

ChatGPT can now buy things for you. On Friday OpenAI launched ChatGPT Wallet, so its AI agents can check out with your card. What’s actually in it: → Your card number isn’t shared with the agent or the store. It pays with a substitute → You review the items and the total, then approve every purchase with Face ID or a fingerprint → No spending limits yet. OpenAI says they’re coming → For now it’s a limited rollout to some US ChatGPT Pro users That approval tap is the part to take seriously. In a Cisco and Carnegie Mellon study, AI agents picked pricier options for users they thought were rich, and a clear budget nearly closed the gap for most models. So if you try it, give the agent a number and read the total before you approve. Would you let an AI agent buy something for you? Sources: OpenAI, TechCrunch #AI# #ChatGPT#

DeepSeek 官方说 deepseek-flash 能直接看图，三种传图方式，一次最多 600 张 建议收藏，用到时照着做。做完你能让 DeepSeek 读截图、看图表、描述图片。 这是 DeepSeek 官方 API 文档 Vision（视觉）一页。先说结论： 官方写明，deepseek-flash 模型可以文字加图片一起输入，用来描述图片、读截图里的字、分析图表。 先说清词。 Vision，视觉输入，意思是让模型看图。 base64，是把图片文件转成一长串文字，直接塞进请求里的编码方式。 file_id，是先把图片传到官方的 Files API（文件接口），拿回来的一个编号，之后用编号代替图片。 先把话说全： 官方支持的图片格式是 JPEG、PNG、GIF、WebP，官方说格式按文件真实内容判断，不看文件名。 官方写明，assistant（模型自己的）消息里放图片，会返回 400 错误；图片只能放在 user、system、tool 消息里。 一、三种传图方式 第一种：base64 内联。官方说最适合本地文件，把图片编码后做成 data: 开头的地址放进请求，整个请求体上限是 48 MiB。 第二种：外部图片链接。给一个公开的 http 或 https 地址，官方限制：链接最长 8192 个字符，图片最大 32 MiB，下载要在 60 秒内完成。 第三种：Files API 的 file_id。官方说同一张图要反复用，或者图太大，就先上传再用编号。这种方式图片可以到 64 MiB。 二、请求怎么写 第一步：model 写 deepseek-flash。官方说旧名字 deepseek-v4-flash-vision-exp 还能用，但已经退役，实际由最新的 Flash 模型处理。 第二步：messages 里 content 不再是一句话，而是一个列表，里面一块是 text 文字，一块是 image_url 图片。 第三步：Responses API 也能传图，图片放在 input_image 里，文字放在 input_text 里。 三、detail 清晰度开关 官方写明，image_url 可以加 detail 字段。low 会把图缩到 512×512 再看，更快更便宜；high、original、auto 目前都是保持原图。 四、费用和上限 官方说图片按尺寸换算成 token，和文字一起计费。每张图最多 1024 个 token，因为太大的图会先被缩到约 1300×1300 的像素量。 官方给的上限：每次请求最多 600 张图；单边最长 8192 像素，一次放 15 张以上时降到 4096 像素；不含 file_id 的图合计 64 MiB，含 file_id 最多 200 MiB。 五、用 Anthropic 接口也行 官方说把 base_url 写成 image 块，source 的 type 可以是 base64、url 或 file。用 file 时要加请求头 anthropic-beta: files-api-2025-04-14。 常见坑 3 条： 一、把图片放进 assistant 消息，会直接报 400。二、外链图片超过 32 MiB 或 60 秒下不完，就改用 file_id。三、一次放 15 张以上，单边别超过 4096 像素。 复制即用的提示词：请描述这张图片，列出图里所有文字，再用三句话总结它想说明什么。 收藏版：5 步速查 一、model 写 deepseek-flash。二、选 base64、链接或 file_id 传图。三、content 写成文字块加图片块。四、格式用 JPEG、PNG、GIF、WebP。五、别把图放进 assistant 消息。 S1哥 观点：官方把每张图封顶 1024 个 token，说明传图比很多人想的便宜，截图问答可以放心大量用。 一句话结论：DeepSeek 读图，认准 deepseek-flash，小图用 base64，大图或重复用 file_id。想让朋友少踩坑，复制链接发给他。 你是想让它读截图还是看图表？卡在第几步？同系列前几篇在我主页。 参考：DeepSeek 官方文档 Vision（S1哥 整理） 今天送：DeepSeek 看图请求自检清单，评论区第一条直接拿。

this paper should scare every agent framework out there researchers took apart 11 coding agents straight from the source code claude code codex gemini cli openhands aider and more roughly 4 million lines not one of them imports a langchain style framework not one of them searches code with embeddings 9 out of 11 already ship skill.md then a second paper went further and let a loop write the harness the model stays frozen a debugger agent reads every failed run an evolve agent edits the harness files the next round keeps the edit or rolls it back 10 rounds later a one tool seed went from 69.7% to 77.0% on terminal bench 2 the codex harness people built by hand sits at 71.9% craziest part editing only the system prompt made it worse all the gains came from memory, tools and middleware agent = model + harness the model takes the bow the harness does the work drop claude fable 5.2 in that frozen seat and the loop rewrites everything around it both papers in the reply

Anthropic set November 19 for its Nasdaq debut, at a $70 to $78 range that values the company at up to $1.45 trillion. In the clip below, CEO Dario Amodei tells NBC News what that means for someone just starting out: "It can do entry-level law work. It can do entry-level finance. It can do entry-level consulting." He also said we can't deny it will eliminate jobs, probably a large number of them. His first advice: learn to use AI. The scarce skill is moving from landing that first job to running the tools that already do it. Is "learn to use AI" still enough, or is that first job already the model's?

Open source models on top again. Deepnote tested 12 models in 50 professional situations where an agent must independently decide whether and where to speak up. DeepSeek V4.1 Flash scored 6.35 out of 10, surpassing Claude Haiku 5.5 (6.05), Gemini 4 Argon (5.60) and Grok 4.7 (4.51). GPT-6 Astra takes first place with a score of 8.97.