Devin Lai

🚀 Building AI systems for the future of humanity
🤖 Autonomous agents, automation & products with real-world impact
📈 Build in public | Startups, wealth creat

[Devin Lai @Devin_Lai_X]

中文：把 FLUX.3 Image Klein 搬到 Mac 本地了。

支持文生图、参考图编辑，提供 5 种模型配置。M5 / 32 GiB 已验证，欢迎更多 M 系列用户实测。

回复你的芯片、内存和 preset，一起补全 Mac 跑分地图。

代码、模型和复现步骤：
https://t.co/Y7HJoSGvW9

[Devin Lai @Devin_Lai_X]

Your Mac. Your AI image studio.

I released FLUX.3 Image Klein for Apple Silicon: local generation + editing with MLX/Metal.

M5 / 32 GiB, 16 steps: 4.16× faster vs my strict MLX FP32 baseline, 43% lower peak memory.

Actual outputs 👇
https://t.co/Y7HJoSGvW9
#MLX

[Devin Lai @Devin_Lai_X]

End-to-end timings on M5 / 32 GiB (Q8 quality, 16 steps):

768×512 teapot: 25.47s
1024×1024 photo: 66.61s

120 fresh-process runs, 4 workloads. Prompts, seeds, original PNGs + raw timings are public.

5 ready-to-run model variants. No conversion needed.

[Devin Lai @Devin_Lai_X]

Claude Code / Codex 额度还剩多少？上下文被什么占了？

试试 Gantry 🐙 原生 macOS 控制台：
• 查看额度与重置时间
• 拆解启动上下文、排查 MCP
• 配置修改可预览、可撤销

公开预览版，个人非商业使用免费。欢迎试用、反馈和 Star！
https://t.co/OZfDzTsAIX

[Devin Lai @Devin_Lai_X]

Meet Gantry — your AI coding setup in one view.

A native macOS dashboard for Claude Code & Codex: cached quota, startup context, MCP diagnostics, and fixes you can preview and undo.

Free for personal, non-commercial use. Closed-source preview.

https://t.co/OZfDzTsAIX

[Devin Lai @Devin_Lai_X]

注册成功后，前往设置页面兑换邀请码。填写好友邀请码，双方均可领取 10 亿 token。
👉邀请码：JF3NZA

注意：邀请码需要注册完成 48 小时内填写，超时无法领取奖励。

[Devin Lai @Devin_Lai_X]

@xueyu1125 有5090八卡机器的可以白嫖部署方��:https://t.co/3AInWijoBo

[Devin Lai @Devin_Lai_X]

@deepseek_ai If you're trying to self-host V4.1-Flash, I put together a vLLM deployment guide with CUDA graph fixes and CPU offload. Quickstart: https://t.co/8R10LIBNnM

[Devin Lai @Devin_Lai_X]

@sairahul1 If you're trying to self-host V4.1-Flash, I put together a vLLM deployment guide with CUDA graph fixes and CPU offload.
Quickstart: https://t.co/8R10LIBNnM

[Devin Lai @Devin_Lai_X]

DeepSeek-V4.1-Flash. 8× RTX 5090. 5.6× output throughput.
6.0 → 33.7 tok/s with CUDA graphs vs patched eager mode (short synthetic test, concurrency 1).
Code + benchmarks:
https://t.co/3AInWijoBo
@deepseek_ai @vllm_project #LocalAI #OpenSource

[Devin Lai @Devin_Lai_X]

DeepSeek-V4.1-Flash on 8× RTX 5090
6.0 → 33.7 output tok/s: 5.6× faster with CUDA graphs vs patched eager mode in a short synthetic benchmark at concurrency 1
Deployment scripts, patches & reproducible benchmarks
https://t.co/3AInWijoBo

[Devin Lai @Devin_Lai_X]

🔥 Selling a Kimi AI Pro account!
💡 Includes full 1-year/monthly tier access (originally ¥199 RMB / ~$28+ value).
💵Price: **$30**
DM me if you’re interested! First come, first served. 📥

[Devin Lai @Devin_Lai_X]

Same prompt. Two models. Two raw, unedited outputs.
One is from MAI-Image-2.5-Pro. The other is from GPT-Image-2. We’ll reveal which is which later.
Zoom in. Pick A or B—and tell me the tiny detail that decided it.
Vote before reading the replies 👇

[Devin Lai @Devin_Lai_X]

Just open-sourced LocateAnything CoreML
Run LocateAnything-3B locally on macOS with CoreML acceleration.@nvidia
Code: https://t.co/44IQWA6JUz
Model: https://t.co/0MTYlI8l0A
Feedback and PRs welcome
#LocateAnything #LocateAnything3B

[Devin Lai @Devin_Lai_X]

Watching GPT roast celebrity styling choices is becoming a genre of its own 😱

Just dropped the Elon Musk Aesthetics Upgrade Report — a full redesign of the hair and wardrobe.

From everyday founder mode to polished future-CEO energy, this is the kind of switch that turns “tech bro” into “tech icon.”

Heavy bangs and awkward cuts had to go. Cleaner lines, sharper fits, better proportions.

Prompt thread below + bonus full outfit breakdown 👇
Who should get the next makeover?
#AIart #ElonMusk #StyleUpgrade #FashionTech #GlowUp #DesignCulture

[Devin Lai @Devin_Lai_X]

@star_okx 在OKX买Euler K3理财，结果1055 USDT被套死无法赎回。官方不作为，客服拿免责条款当挡箭牌。老板能帮忙处理一下吗?至少得催促项目方增加流动性啊!

[Devin Lai @Devin_Lai_X]

@ki_young_ju 92641