lengsh

[AstroHan @AstroHanRay]

如果要用 GLM-5.2， Ollama Cloud Pro 个人测试下面是市面上的最佳选择了，量大管饱，20 美金非多并发基本上用不完。

同价位下的 https://t.co/77Gy4q0tbu 海外套餐性价比完全比不过，而且还有北京时间 14:00-18:00 的三倍消耗，会比较影响效率。

另外，GLM-5.2 是个好模型！

[AstroHan @AstroHanRay]

对比了下，Ollama Cloud 的 tokens/s 比 https://t.co/77Gy4q0tbu 的 Max 套餐高 40%，首字延迟更是低非常多，流式输出几乎秒回。

各家套餐都用下来，我还是觉得 Ollama Cloud 是使用 GLM-5.2 的最佳选择，吞吐快，延迟低，用量大。

[Mia @MiaAI_lab]

Run Gemma-4-26B-A4B on your DGX Spark
256k context • MTP • up to 18 concurrent sessions 🚀

Performance:
1 session → 30 tok/s
2 sessions → 63 tok/s
4 sessions → 109 tok/s
8 sessions → 182 tok/s
18 sessions → ~272 tok/s

Get the full recipe here:
https://t.co/ONKN6ljipR

[Tom Dörr @tom_doerr]

Automates cybersecurity threat detection and monitoring with AI

https://t.co/evOGC7Icsc

[Tom Dörr @tom_doerr]

Aggregated IP blocklists updated every 2 hours

https://t.co/qOHL7nLhGL

[Vivek | Cybersecurity @VivekIntel]

EvilWAF - Advanced WAF Testing & Security Research Toolkit 💀🔥

EvilWAF is an open-source MITM proxy and WAF security testing tool built for authorized penetration testing. It helps security professionals analyze WAF behavior, identify misconfigurations, and evaluate defensive mechanisms without modifying payloads from testing tools.

✨ Key Features:
• 🛡️ Transparent MITM proxy for security testing
• 🔍 Deep multi-layer WAF vulnerability scanning
• 🔄 TCP, TLS & HTTP/2 fingerprint rotation
• 🌐 Origin IP discovery behind protected websites
• 🚀 Works with sqlmap, ffuf, nuclei & other proxy-supported tools
• 📊 Real-time TUI dashboard with detailed scan insights
• 🐍 Built with Python and Docker support

🔗 https://t.co/IAuNMk4nE0

#CyberSecurity #Pentesting #AppSec #Python #OpenSource #GitHub #InfoSec

[Vivek | Cybersecurity @VivekIntel]

Security Onion - Open Source Platform for Threat Hunting & Security Monitoring

Security Onion is a powerful Linux distribution built for SOC analysts, threat hunters, and blue teams. It combines industry-leading security tools into a single platform for network monitoring, intrusion detection, log management, and incident investigation.

✨ Key Features:
• 🛡️ Enterprise Security Monitoring & Threat Hunting
• 🚨 Network IDS with Suricata & Host Monitoring
• 📡 Network Metadata powered by Zeek
• 📦 Full Packet Capture (PCAP) for deep investigations
• 🔍 Elastic Stack for powerful search & analytics
• 🖥️ Unified SOC dashboard for alerts, hunting & case management
• ☁️ Supports AWS, Azure & Google Cloud deployments

🔗 https://t.co/iLu8bPzmbq

#CyberSecurity #ThreatHunting #BlueTeam #SOC #OpenSource #GitHub #NetworkSecurity

[Viking @vikingmute]

看到别人发的 CF 的这个开源 Agentic Inbox 挺有意思的，https://t.co/L7vVJJSX9S

自托管邮件客户端 + AI agent 回复，里面用的都是 CF 自家的服务，Email Routing 接受，Email Sending 发送， Durable Objects + D1 存储，R2 静态文件。
Cloudflare Workers AI 和 Agents SDK 内置 agent sdk，帮你恢复邮件。

我觉得针对我这种好多个域名都在 CF 的，并且每个域名都需要接入支持的 email 地址的确实挺不错的，现在我还在忍受着不太好用的 Zoho 企业邮箱，不知道有人试过没有。

��是现在大公司的通病就是搞开源像秀肌肉，这个项目发布以后就不更新了，3 个月没更新，issues 躺一大堆的东西也没人管了。好多 Vercel 的项目也是一样

[Hugging Models @HuggingModels]

Ever wished for an AI that thinks, codes, and creates without limits? Meet Qwen3.6-40B: a multi-stage tuned, uncensored giant built for image-text-to-text tasks. It's abliterated for raw creativity, and it's blowing up with 504k downloads. This is not your average model.

[Tom Dörr @tom_doerr]

Indexes codebases into five intelligence layers

https://t.co/1Y6FD1JOIw

[Yero @Yerocode0]

🚨 ¡BOMBA!

China acaba de lanzar un agente de IA que automatiza tu escritorio 100% en local.
Sin internet. Sin nube. Sin mandar tus datos a nadie.

Puede:
✅ Controlar cualquier aplicación de escritorio
✅ Abrir archivos y carpetas
✅ Navegar por webs
✅ Hacer tareas completas de forma autónoma

Y lo mejor: es 100% código abierto.
Esto ya no es futuro… es ahora.

¿Cuál sería la primera tarea que le pedirías? 👇
RT si te flipa y sígueme para más IA real que importa

[divyansh tiwari @DivyanshT91162]

This might be the most underrated MCP server I've found this week.

Stealth Browser MCP gives AI agents an undetectable browser that can automate sites most browser tools can't.

• Bypasses Cloudflare & anti-bot systems
• 97 browser automation tools
• AI-generated network hooks
• Pixel-perfect UI extraction
• Full CDP access
• Cross-platform support

100% Open Source (MIT License)

Repo:
https://t.co/66pVMNWdok

[Md Ismail Šojal 🕷️ @0x0SojalSec]

Voice cloning on CPU locally run regular lowend laptop,

This TTS a blazing-fast, lightweight text-to-speech model that runs smoothly on regular CPUs.

- 200ms first chunk
- Voice cloning
- Multiple languages support
- 6x real-time on a CPU

- https://t.co/ipT9Z2ork3

[徐伦 @Jolyne_AI]

开源 AI 网页自动化工具：Nanobrowser。

OpenAI Operator 的开源替代方案，本地在浏览器里运行，支持多智能体协作。

免费、重视隐私、LLM 选择灵活、代码完全开源，让网页操作更智能、更高效。

GitHub：https://t.co/8q3Pl4eECO

从 GitHub 下载 zip，导入 Chrome 扩展即可开始使用。

[GitHub Projects Community @GithubProjects]

MaxKB is an open-source platform for building enterprise-grade RAG agents with an integrated pipeline for knowledge retrieval and AI workflows.

- RAG pipeline with automatic text splitting, vectorization, and document ingestion
- Agentic workflow engine with function library and MCP tool-use capabilities
- Model-agnostic support for private (DeepSeek, Llama, Qwen) and public (OpenAI, Claude) models
- Zero-coding integration into third-party business systems

[Tech2Wild @Tech2Wild]

If You Want To SEE How Tok/s Actually Looks. Like 5,10,20,30 all the way to 2000tok/s I found this dope site that helps you visualize it. I can say even at 20 tok/s speeds are not that bad especially for on a conversational scale. https://t.co/v8rRrWA7fk

[Guillermo Casaus @_guillecasaus]

No vuelvas a usar Ollama si quieres sacar el máximo rendimiento a tu IA local.

No exprime al máximo la potencia de tu GPU.

Y eso se nota cuando empiezas a servir modelos más exigentes.

Por eso nació vLLM.

Un proyecto desarrollado originalmente en UC Berkeley para ejecutar modelos de IA de forma mucho más rápida, eficiente y preparada para producción.

→ Hasta 2 veces más rendimiento en determinadas cargas de trabajo
→ Compatible con más de 200 arquitecturas de modelos
→ API compatible con OpenAI para integrarlo fácilmente
→ Funciona con NVIDIA, AMD, Apple Silicon, TPU y mucho más

Se instala con un simple:

pip install vllm

Mientras Ollama está pensado para que cualquiera pueda ejecutar modelos de IA...

vLLM está diseñado para exprimir al máximo el hardware y servir modelos con el mejor rendimiento posible.

Tiene más de 85k stars en GitHub, es 100% gratis y open-source.

Repo en comentarios 👇

[huangserva @servasyy_ai]

如果你一个人维护项目，没有安全团队，上线前心里总没底，那么这东西基本就是给你的礼物

说实话，市面上叫"AI 安全扫描"的东西我基本不信。

扫完给你报一屏警告，九成是误报，纯浪费时间。

直到我扒了下这两天 GitHub 日榜第一的 strix，它不一样。

它不扫描，它是直接黑进你代码里。

在 Docker 沙箱里把应用跑起来，带一个 Caido 的 HTTP 拦截代理改包，开浏览器打 XSS、越权，还有个 Python 沙箱专门写 PoC。

区别在哪？

传统工具告诉你"这儿可能有 SQL 注入"。
而strix 直接写一段能打通的 exploit 给你，附复现步骤 + CVSS 评分。

SSRF、反序列化 RCE、JWT、业务逻辑里的竞态和支付篡改，OWASP Top10 往外全在打。
并且多个 agent 还能组队分工。

打完自动生成补丁开成 PR，能挂 CI/CD 每个提交拦一遍。

以前请人做一次渗透测试，几周起，几万块；现在几小时，一个 key 的钱。

Apache 2.0，配个 LLM key 就跑👇

https://t.co/wVgK0o4hAZ

[AJ @ItsmeAjayKV]

It's crazy to think that most of my tasks can be done with local models now.

My current local model stack for @NousResearch Hermes on my single 3090:

Gemma4-12B-QAT (KV F16 + full context)-> normal everyday tasks. quick research.

Qwen3.6-35B-A3B-Q6/Q8 (KV q8_0 + 144k context + offloading) -> serious agentic tasks + coding tasks

Qwen3.6-27B-Q4/Q5 (KV q8_0 + 110k context) -> only for coding tasks

All running through llama.cpp.

[Tech2Wild @Tech2Wild]

🚀 Updated my DeepSeek-V4-Flash-DSpark recipe for 2× DGX Spark (GB10)

⚡ Cold-start garble → fixed at the ROOT: a scheduler-level patch (Patch 3, h/t roady001), not a config band-aid
🔧 DSpark spec bumped 3→5 now that it's garble-safe
📊 Honest speed: ~50-60 tok/s on structured/agentic work (JSON, code, tool-calls), ~49 mixed. Deterministic best-case touches ~75 🤷 — but that's cherry-picked, spec-decode speed tracks the workload

NVFP4-KV · TP2 · 1M ctx · vLLM
Full recipe + benchmarks 👇
https://t.co/nCHBmdFEyP