| Website | https://goarcherdynamics.com |
Tested DeepSeek R1 8B Q8 on my Linux server.
48 tokens/sec. 9.6 GB VRAM. Correct code, solid reasoning.
Paste the same prompt 4 times and watch it spiral into identity crisis while writing novels!
It spent 92 seconds debating whether 3.0 is a prime number. Yep!
Read more fun below.
Meet Astra. That's the name Mistral Nemo 12B chose for herself — poetic, self-aware, just like her 7B sibling Elysium. 47 tokens/sec. 9.2 GB VRAM. Best explained code in all my tests. Response got richer and more thorough across the runs -- unprompted. This is the patient teacher you reach for when you need to understand the answer, not just receive it.
Europeans are doing something right. And they're keeping it open source.
Full breakdown below.
Tested Google's Gemma3 12B QAT on my home Linux server. Stable 97% GPU utilization, no CPU spill, no logic errors. Mistral Nemo 12B beats it on speed & uses 2 GB less VRAM. Those extra 2 gig could run a second model on a 16GB card.
Gemma 12B is correct, thorough and about as warm as a DMV waiting room.
Full breakdown below.
