#aiathome — Public Fediverse posts

I said "hey." One word. Three hours after my last benchmark run. DeepSeek R1 8B responded with 1,360 tokens of unprompted Python code — its best output of the entire test series. Then it explained why. And got everything wrong. Perfect recall. Wrong count. Misread my mood. It didn't lose data — it rewrote the narrative.
Turns out the best output comes when you ask for nothing.

Full breakdown below. 👇

#AIatHome #LocalLLM #DeepSeek #Ollama #HomeLab #AI #MachineLearning

https://goarcherdynamics.com/2026/03/23/deepseek-r1-8b-lost-in-time/?utm_source=mastodon&utm_medium=jetpack_social

Tested Mistral's Ministral 3 14B on my home rig — and it has opinions.
Typo-corrected me on first contact. Just sass, straight out of the box.
Performance? Solid 48 t/s, 10GB VRAM, 98% GPU utilization.
Oh, and it can see now. OCR + image recognition.
It's now my daily driver.

Full review on the blog. 👇

#AIatHome #LocalLLM #Mistral #Ollama #HomeLab #RTX5060Ti #MachineLearning #AI

https://goarcherdynamics.com/2026/03/20/ministral-3-14b-can-now-see/?utm_source=mastodon&utm_medium=jetpack_social

Tested Qwen2.5 Coder 14B on my Linux server. Same code, same prompt I used on DeepSeek R1 last week. 12 seconds. Consistent across 3 runs. Zero drama. It cleaned up the code nicely and didn't go insane.
Two models + same task = completely different personalities.
Now I understand why many devs use Qwen Coder as their tool of choice.

Read below to find out how it can help you.

#LocalAI #Qwen #CodeRefactoring #AIAtHome #Ope

https://goarcherdynamics.com/2026/03/19/aihome-qwen-2-5-coder-14b-refactoring/?utm_source=mastodon&utm_medium=jetpack_social

lama déclenche l'apocalypse
https://github.com/ggerganov/llama.cpp

#ml #ai #aiathome