HackerTrans
TopNewTrendsCommentsPastAskShowJobs

kekePower

no profile record

Submissions

[untitled]

1 points·by kekePower·8 месяцев назад·0 comments

[untitled]

1 points·by kekePower·9 месяцев назад·0 comments

[untitled]

1 points·by kekePower·9 месяцев назад·0 comments

[untitled]

1 points·by kekePower·в прошлом году·0 comments

[untitled]

1 points·by kekePower·в прошлом году·0 comments

comments

kekePower
·в прошлом году·discuss
Just pushed a major update: a full translation layer is now live. You can serve any HTML file and automatically translate it on the fly with a local LLM. No client JS, no cloud, no bullshit. Looking for testers, bug hunters, and feedback from anyone running Apache or interested in offline-first web tech. Supports 46 languages with reliable LLM coverage (curated for quality, not just quantity).
kekePower
·в прошлом году·discuss
I’ve been building this in my spare time over the past few days. mod_muse-ai is an Apache module written in C that lets you serve HTML pages generated by an AI model (local or remote). It reads `.ai` files, merges in a system prompt and a layout prompt, and sends the request to either a local Ollama instance or any OpenAI API-compatible endpoint (like LM Studio, Mistral, etc.), then streams the HTML response back to the browser.

The goal is to eliminate the need for a separate backend for AI-generated content. All feedback welcome—especially on the architecture and config design.
kekePower
·в прошлом году·discuss
It offloads to system memory, but since there are "only" 3 Billion active parameters, it works surprisingly well. I've been able to run models that are up to 29GB in size, albeit very, very slow on my system with 32GB RAM.
kekePower
·в прошлом году·discuss
I have an RTX 3070 with 8GB VRAM and for me Qwen3:30B-A3B is fast enough. It's not lightning fast, but more than adequate if you have a _little_ patience.

I've found that Qwen3 is generally really good at following instructions and you can also very easily turn on or off the reasoning by adding "/no_think" in the prompt to turn it off.

The reason Qwen3:30B works so well is because it's a MoE. I have tested the 14B model and it's noticeably slower because it's a dense model.