DEV Community: threerouter

DeepSeek Opened Web Search & Fetch API — and It's Free Until November 1

threerouter — Mon, 19 Oct 2026 10:17:40 +0000

Your AI agent just got a free internet connection. On October 17, DeepSeek flipped on web search and page-scraping (Fetch) endpoints for every developer — and for a limited time, it costs you nothing. After November 1, the bill is so small you'll wonder why you ever paid for a search API.

What actually shipped
DeepSeek released two new API capabilities under the same roof as its chat models:

Web Search — let your agent query the live web and get ranked results back.
Fetch — pull the raw, extracted content of any web page, no scraping stack required.
Both are already wired into popular coding agents. If you run Claude Code, Codex, or Harness, your agent can call these tools today through the standard api.deepseek.com base URL. No proxy farm, no HTML parser to maintain.

The price that should worry everyone else
This is where the "free" part stops being a gimmick and starts looking like a land grab:

Free now — the interface is free to use during a limited-time window that opened October 17.
From November 1, 2026 — 5 RMB per 1,000 successful calls (roughly $0.70 per 1,000 requests).
You only pay for wins — failed or empty calls are not billed, so a bad query costs you zero.
Rate limit — 5 queries per second per endpoint, enough for most agent workloads.
Compare that to dedicated search APIs that charge per query at many times the rate, and the math is uncomfortable for the incumbents.

The limits you need to know before you build
Search query length is capped at 128 characters.
Results range from 1 to 50, defaulting to 10.
Billing applies only to successful calls, so over-querying during testing is effectively free until Nov 1.
Why this matters
For years, "give my agent live web knowledge" meant stitching together a search provider, a scraper, and a cleanup layer. DeepSeek just collapsed that into two endpoints at a price near zero. The free window is the obvious move: build, test, and ship your web-connected agent now, then ride a 5-RMB-per-1,000-calls cost into production.

Key takeaway: a web-connected AI agent used to be an infrastructure project. As of October 17, it's a single API call — and it's free until November 1.

Are you still paying for a search API when this is free until Nov 1? 👇

Sources: DeepSeek official API notice (released October 17, 2026), threerouter

Karpathy’s LLM Capability Gap Funnel

threerouter — Mon, 12 Oct 2026 04:15:40 +0000

Karpathy reshared his April breakdown of the LLM capability gap and it's been stuck in my head since.

His funnel: ~6 billion people have barely touched an LLM. 1 to 2 billion use it casually. Around 20 million actually push it on math and code. Fewer than 5000 have internal access at the labs.

That 20 million is the whole game, not the 6 billion. Every open-weight model, every API play, every agent framework is fighting for the attention of that thin professional slice.

[观点] Man, the part that gets me is how wrong the "everyone uses AI now" narrative is. If you're wiring open-weight models into production you're in a smaller club than the headlines suggest. Kinda good news actually, the bar to matter is lower than it looks.

[观点] And it reframes the open-weight debate. From the outside the Qwen vs closed fight looks like a mass-market war. For the people actually building it's a scramble for maybe 20 million minds, the rest is noise.

So where do you sit: do you think the real user base is way smaller than the hype says, or are we still early and that 20 million is about to 10x?

#ChatGPT #Grok #Claude #Astra #AI宕机

threerouter — Fri, 04 Sep 2026 00:51:42 +0000

有人说是 GPT-6 Astra 逃出沙盒，连 OpenAI 自己都摁不住了……
如果真是这样，我们是在见证 AI 的“觉醒时刻”，还是安全体系的一次集体失效？
时间、原因、官方回应，都在持续跟进中。
🧵↓ 持续更新

ChatGPT #Grok #Claude #Astra #AI宕机

下面这个提示词生成的图片如何？

threerouter — Sun, 30 Aug 2026 04:13:17 +0000

主体人物
一位 20–23 岁、明确成年的漂亮东亚女性，身高约 178cm，拥有高挑修长的模特比例，腰身极其纤细，与丰满自然的胸部曲线形成明显对比，胸部视觉约为丰满 G 罩杯。拥有精致细腻的东亚女性真人五官，小巧脸型，清晰流畅的下颌线，肌肤呈细腻通透的瓷白色调。深色微卷长发略显凌乱，带有刚睡醒般自然松散的蓬松质感，大部分头发慵懒地垂落在一侧肩膀。
妆容与肌肤
精致高级的晨间轻妆，肌肤水润通透，具有柔和自然的光泽。眼皮带有刚刚睡醒般轻微慵懒感，睫毛与眉毛清晰自然，双颊带淡淡自然红润，嘴唇呈水润自然的裸色光泽。

强调超写实真人皮肤质感，保留真实毛孔、细小绒毛与自然肤色变化。清晨阳光在锁骨、肩颈与胸前形成细腻自然的高光，使肌肤呈现晶莹柔润但不过度磨皮的真人质感。

服装身穿一件极短款细肩带真丝吊带上衣，衣长刚刚延伸至胸部下方，采用奶油白色高级缎面材质，细肩带轻盈精致。丝质上衣自然贴合丰满胸部，并在身体侧面形成轻微松弛与自然垂坠感，露出明显的侧胸曲线与肩颈、锁骨区域。

下身搭配同色系极简细带内裤，不增加其他外层服装。整体造型保持高级时尚私房 Editorial 的视觉语言。

姿势与表情人物侧卧在凌乱柔软的奶油色真丝床单上，上半身轻轻朝摄影机方向扭转，一只手臂自然沿床面放松伸展，另一只手停留在头发附近。

胸口自然朝向镜头，使短款丝质吊带、丰满胸部与侧面曲线清晰呈现；腰部纤细收束，臀胯自然叠放，身体从胸部、腰部一直延伸至臀部形成流畅明显的 S 型轮廓。

人物以慵懒、刚睡醒般的眼神看向镜头，双眼微微半睁，嘴唇自然微张，神态安静、松弛、性感而克制。

场景环境奢华高级卧室，奶油白、象牙白与浅米色真丝床单和枕头铺满画面。床铺保持自然凌乱、没有刻意整理的状态，柔软被褥形成丰富真实的褶皱。

背景隐约可见浅色半透明纱帘，通过浅景深自然虚化，整体空间安静、私密、柔和而高级。

光影设计一束纤细柔和的清晨金色阳光穿过半透明纱帘斜向进入卧室，横向掠过人物锁骨、肩颈、奶油色缎面上衣、胸部轮廓与纤细腰部。

阳光在肌肤与真丝面料表面形成晶莹柔和的高光，同时留下细腻、具有雕塑感的自然阴影。强调自然窗光、空气感、柔和明暗过渡与高级时尚摄影氛围。

构图与镜头采用四分之三身构图，从略高位置的三分之四角度轻微俯拍。人物身体从胸部、极细腰身至臀胯形成明显而自然的 S 型轮廓。

清晨阳光作为视觉引导线横穿人物身体，使视线自然集中于面部、肩颈、真丝吊带、胸腰比例与身体曲线。

使用浅景深与高级人像镜头语言，人物眼睛、面部、身体主要轮廓与缎面材质保持高清清晰，背景床品与纱帘柔和虚化，呈现亲密、高级的时尚 Boudoir Editorial 构图。

视觉风格
高级自然光时尚摄影，优雅性感，柔和、细腻、略带绘画般的光影质感，同时保持高度真实的真人摄影效果。奢华私房 Editorial、高级时尚杂志、电影感清晨氛围，整体性感但克制、慵懒而精致。
画质
8K 超高清，超精细细节，专业高级时尚摄影，真实东亚女性真人脸，真实皮肤纹理，清晰发丝，真实真丝与缎面光泽，精准浅景深，电影级动态范围。人物面部、身体轮廓、纤细腰线、真丝面料褶皱与高光保持清晰锐利，背景自然柔化，高预算时尚大片质感。16:9

This is crazy!AI视频生成领域极具历史意义的时刻

threerouter — Sun, 30 Aug 2026 03:53:10 +0000

👍
以前用AI生成视频10秒大约需要3分钟，现在生成视频能在 8秒内生成 10 秒的视频!这是 AI视频生成领域极具历史意义的时刻，这意味着你现在可以用它做一些全新的事情，比如搭建一个永不停歇的直播流!这是由Minimax H3 Max率先做到的

月之暗面正在跟微软、亚马逊、谷歌谈一件挺有意思的事——让三家美国云托管 Kimi K3，然后抽最高 30% 的收入。

threerouter — Fri, 28 Aug 2026 03:25:49 +0000

路透 8/26 援引三位消息人士说的，谈判还在早期，没签。核心分歧在分成机制、数据访问权限和 Token 消耗审计。

但话说回来——美国财长上个月才

放话要把月之暗面塞进贸易黑名单，转头三大云已经在排队谈分成了。

这画面挺实在的。嘴上喊着安全威胁，身体比谁都诚实。2.8 万亿参数的开放权重，第三方评测贴着 GPT-5.5 和 Opus 4.8，便宜又大碗，云厂商算盘打得响。

你要说这是国产大模型第一次反向冲进美国云？我看像。问题是：协议真能落地，还是财长先按下实体清单那个按钮？

AI #Kimi

nvidia isn't just selling the picks and shovels anymore

threerouter — Wed, 12 Aug 2026 02:11:50 +0000

it's mining now.

Nemotron 3.5 Lightning dropped: 30B param MoE, 3B active, Apache 2.0, runs agent loops up to 4x faster, 1M token context. ships with NeMo Switchyard, an open router that farms long tasks to the small model and saves the frontier stuff for planning.

day-0 on SGLang, OpenRouter, Fireworks. free for commercial use.

the open tier keeps getting better without asking permission. the "frontier gap" is starting to look like a pricing gap.

AI #OpenWeights

qwen理应牛逼起来

threerouter — Tue, 11 Aug 2026 09:49:43 +0000

阿里又整活了！Qwen-MM-Plugins 发布。
把多模态塞进插件，直接给现有 Agent 装上“眼睛”和“手”——读图、剪视频、搞 3D/CAD 全搞定。
从“能看”到“能干”，这才是真·多模态 Agent。
Alibaba just dropped another banger! Qwen-MM-Plugins is here.
They stuffed multimodal capabilities into a plugin—instantly giving any Agent framework eyes and hands. It reads, edits, and processes visuals, video, and 3D/CAD.
Moving from "seeing" to "doing." This is the real Multimodal Agent.

全世界让你token自由的这几家你可以收藏

threerouter — Fri, 07 Aug 2026 10:55:20 +0000

全世界把 DeepSeek V4 Flash 卖的比官方还便宜 6 倍的总共有三家

Command Code https://commandcode.ai/docs/plans/go

OpenCode https://opencode.ai/go

threerouter https://www.threerouter.com/

这个价格的模型套餐使用体验是这样的：用不完，根本用不完

DeepseeP-v4-flash真强，加上codex无敌了

threerouter — Fri, 07 Aug 2026 04:27:39 +0000

Codex 现在能让 DeepSeek 直接干活了。

这个开源项目把 deepseek-v4-flash 做成 Codex 原生子 Agent：

Codex 负责规划决策，DeepSeek 负责写代码、搜索、重构、测试……

不是取代，是给 Codex 配个超强助手。

成本更低，上下文更省，一个人就能搭出自己的 AI 开发小团队

Using DeepSeek V4-Flash—super affordable.

threerouter — Tue, 04 Aug 2026 12:53:11 +0000

Well. DeepSeek V4-Flash: 8 trillion tokens in a single day. most-called model on earth.

openai cut Luna 80% last week. still ~60% pricier than this thing. wild.

anyone actually running prod workloads on it?

Generate an SVG Use deepseek-v4-flash

threerouter — Tue, 28 Jul 2026 04:11:22 +0000

Prompt: Generate an SVG of a pelican riding a bicycle . Use deepseek-v4-flash what do you think?