---
title: "Dragos Roua (@dragosroua) | Hashnode"
description: "I code since before it was vibey. Read the latest articles by Dragos Roua on Hashnode."
url: "https://hashnode.com/@dragosroua"
---

@dragosroua

# Dragos Roua

@dragosroua · Lisbon, Portugal · Joined September 2026

I code since before it was vibey

[Website](https://dragosroua.com/)

### [\[ Video \] Local AI: Running Alpie Seed on a 16GB MacBook Pro](https://dragosroua.hashnode.dev/video-local-ai-running-alpie-seed-on-a-16gb-macbook-pro?utm_source=hashnode&utm_medium=feed)

The 14th models in my series about local AI that fits on a MacBook Pro was Alpie Seed. A Qwen post-train, usable for moderate daily work.

2 min read · [dragosroua.hashnode.dev](https://dragosroua.hashnode.dev/video-local-ai-running-alpie-seed-on-a-16gb-macbook-pro?utm_source=hashnode&utm_medium=feed)

[10h ago](https://hashnode.com/posts/video-local-ai-running-alpie-seed-on-a-16gb-macbook-pro/6af9616bb31b15dddf1f4935)

### [\[ Video \] Running Nanbeige 4.2 3B on my 16GB MacBook Pro](https://dragosroua.hashnode.dev/video-running-nanbeige-4-2-3b-on-my-16gb-macbook-pro?utm_source=hashnode&utm_medium=feed)

I tested Nanbeige 4.2 3b on a 16GB MacBook Pro. The 4bit quant is good, but the generation is relatively slow

2 min read · [dragosroua.hashnode.dev](https://dragosroua.hashnode.dev/video-running-nanbeige-4-2-3b-on-my-16gb-macbook-pro?utm_source=hashnode&utm_medium=feed)

[5d ago](https://hashnode.com/posts/video-running-nanbeige-4-2-3b-on-my-16gb-macbook-pro/6af2bf34cc1e239b8e7cee78)

### [\[ Video \] Running K2 Horizon 3.7B on my 16GB MacBook Pro](https://dragosroua.hashnode.dev/video-running-k2-horizon-3-7b-on-my-16gb-macbook-pro?utm_source=hashnode&utm_medium=feed)

K2 Horizon 3.7B runs comfortably on a 16GB MacBook Pro. Worth noting that the 7B parameter version was better, but way too slow.

2 min read · [dragosroua.hashnode.dev](https://dragosroua.hashnode.dev/video-running-k2-horizon-3-7b-on-my-16gb-macbook-pro?utm_source=hashnode&utm_medium=feed)

[Nov 8](https://hashnode.com/posts/video-running-k2-horizon-3-7b-on-my-16gb-macbook-pro/6af076cbc8049b8101fee16f)

### [Can AI Improve Itself Without Training? Well, Maybe…](https://dragosroua.hashnode.dev/can-ai-improve-itself-without-training-well-maybe?utm_source=hashnode&utm_medium=feed)

A frontier model was capable to change a smaller model by just editing their weights. The short and long term consequences are fascinating.

3 min read · [dragosroua.hashnode.dev](https://dragosroua.hashnode.dev/can-ai-improve-itself-without-training-well-maybe?utm_source=hashnode&utm_medium=feed)

[Nov 5](https://hashnode.com/posts/can-ai-improve-itself-without-training-well-maybe/6aec8ad84c755372f06d7f0c)

### [Running A Billion Parameter Model (LFM3 1.7B) on a Raspberry Pi](https://dragosroua.hashnode.dev/running-a-billion-parameter-model-lfm3-1-7b-on-a-raspberry-pi?utm_source=hashnode&utm_medium=feed)

A Raspberry Pi 4 can run comfortably (3t/s) LFM3 1.7B. Not usable for heavy coding, but good enough for an orchestrator.

3 min read · [dragosroua.hashnode.dev](https://dragosroua.hashnode.dev/running-a-billion-parameter-model-lfm3-1-7b-on-a-raspberry-pi?utm_source=hashnode&utm_medium=feed)

[2](https://hashnode.com/posts/running-a-billion-parameter-model-lfm3-1-7b-on-a-raspberry-pi/6aeb5f5bc70a133bf2cb160d) [Nov 4](https://hashnode.com/posts/running-a-billion-parameter-model-lfm3-1-7b-on-a-raspberry-pi/6aeb5f5bc70a133bf2cb160d)