Blog

Notes from building a local AI coding setup on a consumer GPU: measured numbers, dead ends, and what comes next.

I'm building a local LLM that actually gets better the longer I use it

Shura: a local AI coding assistant on one RX 7800 XT that should keep improving from the work we do together. The plan, the principles, and what I don't know yet.

Decode with MTP: 78 tok/s
Context: 200k
Prefill, 100k-token input: ~1,000 tok/s

AMD
MoE
lifelong

November 9, 20265 min read

Read the post

About the project

Shura is a local coding assistant that runs on one AMD RX 7800 XT with 16 GB of VRAM. Everything stays on the machine. The code is open on GitHub.

To follow new posts, paste the feed address into any RSS reader.