Andrew Zhu

Today, book "Using Stable Diffusion with Python" is published. Hope it is useful for SD developers. #stablediffusion #diffusion #ai amazon.com/Using-Stable-D…

my dear friend, go build your own AI rig, it is not hard, a bit costy, but you will thank your move

@0xSero @teortaxesTex Everyone should do this, maybe one GPU, or two, but need to do it now, not tomorrow

@teortaxesTex Can I interest you in some self sovereignty

wtf man

@XDash 鸵鸟说, 我啥也没听见, 到底发生了什么? 嗯,好像什么也没发生

我实践了半年，不听任何 AI 创业者高谈阔论的播客节目，无论嘉宾履历多光鲜、播客厂牌多知名。 我错过了什么吗？什么都没错过。

@LottoLabs I will stick with Qwen3.6-27B, it do whatever I told it to do, and do it great untile Qwen3.7-27B, opus 4.8? what is this thing

Pretty crazy number out up by 4.8 but does it pass vibes? Idk if it’ll win me over from 5.5 and open source models Anthropic desperately needs to have big usage plans

@donkey_kongfu @hellojixian 待机就十几瓦

@hellojixian 1200w，一个小时 5 毛，一天 10 块。讲道理，电费比 dstoken 费用高

塞得满满当当。这可能是最便宜的32G显存的AI配置了，跑个31B的模型+262K长上下文 速度其实比我Mac Studio M3 Ultra 还快😓 感觉又花冤枉钱了

@TheAhmadOsman I am building a fast and clean one, and I think everyone should build one, no more bloating, no more token wasting

Agentic harnesses are so bloated right now btw This is the slowest and worst they’ll ever be

@GoSailGlobal 又傻又慢, 没法用, 不过这个设计倒是不错, 值得研究一下

字节悄悄把 GUI Agent 这条路线开源了，而且做得比想象中扎实 UI-TARS-desktop（GitHub 29.4k ⭐）一个仓库里塞了两个东西： · Agent TARS：通用多模态 Agent 框架，CLI 一键启动，能在终端 / 浏览器 / 电脑里跑真实任务（订机票、订酒店、画图都演示过） · UI-TARS Desktop：本地 GUI Agent，看屏幕、点鼠标、敲键盘，全部本地跑，不上传 Apache 2.0、原生支持 MCP，背后是字节自家的 UI-TARS 视觉模型 + Seed-1.5-VL 国产开源 Computer Use 这条线，目前最完整的一份 🔗 github.com/bytedance/UI-T…

Why AI Agent Slow down After Memory Compaction, And How I Fixed It @xhinker/your-ai-agent-freezes-every-20-minutes-heres-why-and-how-i-fixed-it-eb6ef7d524b8" target="_blank" rel="nofollow noopener">medium.com/@xhinker/your-…

@LottoLabs No, they don't know, even Ali employee don't know it either

Does the average Chinese person know they have the best open source models in the world?

@Aditya_181105 To me, it is Qwen3.6-27B

Devs, Which model actually writes production ready code?

@LottoLabs Alibaba's Qwen team is primarily located in Hangzhou, which is the home city of Alibaba's headquarters. The core AI division—Tongyi Lab—is based there. Looking forward to your good news, let us know, I also ASKED my friend in Qwen/Alibaba team to open source it SOON :D

@xhinker I just assumed qwen’s headquarted in Beijing

This isn’t even expensive

@LottoLabs Exactly, and this will be the survival strategy for every individual who want to stay independent and self-reliant

The skills you learn from running local models is more valuable than the cost of the hardware

@0xSero @LottoLabs wow, 200k context, this is amazing

@LottoLabs 27 tok/s achieved

LFG Finally Deepseek-V4-Flash on 1x DGX Spark - No MTP - 12.5 tok/s decode - 555 tok/s prefill - 200k context The whale is officially working perfectly on 1x DGX Spark so far no measured regressions but during testing period if you report issues I will identify and rectify

I just learned one of my previous colleague and friend is working closly with Qwen team, I asked for Qwen3.7-27b release

#Qwen3.6-27B IS BETTER and SMARTER than Deepseek-V4-Flash!

You should buy an RTX 3090 and learn how to run models locally The elite don’t want you to know this but running local models is hella easy, performant, and cheap nowadays

@LottoLabs 100% +1

3.6 was such a solid update Patiently waiting for 3.7

@NoahKingJr Robot

What’s coming after Artificial intelligence?

@Hikari_07_jp You are 100% on the right track

Local LLM is incredibly complex. Hardware selection, quantization, harnesses, engines, tensor parallelism, unmodified models, MTP… Despite its complexity, local LLM is irresistibly fascinating. I started using X because there was almost no one close to me who could share this excitement with me.