[nekusar]

Local LLMs.

Krasis is one such tool that allows large models using blended GPU/RAM.

ik_llama for better performance than llama.

ComfyAI for local image generation.

Nanocrab seems better for orchestration. Still need a good system capability firewall.

[nativeit]

Who’s buying the memory for this effort?

[bad_haircut72]

Think how cheap its gonna be when everyone abandons the cloud providers and they start selling the 50B of hardware they over-invested in

[nekusar]

I got 96GB of DDR5 ram 2y ago for $300.

Which now, 32GB goes for $300. Fucking insane. But prices will eventually come down as the enterprise and corpo scalpers realize AI is a losing deal for human replacement. Nvidia has already said as much.

https://fortune.com/2026/04/28/nvidia-executive-cost-of-ai-i...