Tested FlappyBench on Qwen 3.8 27B, DeepSeek V4 Pro 0813, and Gemini 3.7 Flash with /design command.

All scored 9/10, it comes down to cost, quality, and iterations. 🔹 DeepSeek V4 Pro ($0.0036): Low cost, solid, but movement and spacing need work. 🔹 Qwen 3.8 27B ($0.0068):

Command CodeAug 2135

Too many good new tools out there it’s hard not to switch and try them all every week.

But this is what I’ll likely be sticking with. NousResearch Hermes new web UI with the group chats is so good. Plus tailscale letting me watch and interact from my phone. I have grok bossing

jakeAug 20

Showing the true power of the DGX Sparks : The AI Cookshow.

Two Sparks. One night. A model fleet wrote a playable roguelike from a single prompt, and then one of the models sat down and played its own game. No API. Zero cloud tokens. The setup: - Spark 1 runs Qwen 3.8 27B dense

Yume_XAug 2012

I’m testing local llms for task analysis in Levelyss.

Same task: “Gain 500 followers on X” Gemma 4 12B - 45 XP - High importance - Skills: Marketing, Writing, Networking - Category: Work - 3.4s Gemma 4 26B - 35 XP - Medium importance - Skills: Marketing, Communication,

EvgeniiaAug 201