Lo Merlins @LoMerlins

Local LLMs, AI agents & Code testing github.com/LoMerlins/agen… Joined April 2023

Tweets

27
Followers

136
Following

52
Likes

95

Lo Merlins @LoMerlins

2 months ago

@0xCapexOG For many people the privacy + independence is worth the small performance gap. But I agree, free isnt free of tradeoffs.

0 0 2 19 0

View Details

Lo Merlins @LoMerlins

2 months ago

@0xCapexOG Appreciate! More good stuff coming

0 0 1 26 0

View Details

I'm building a Diablo 2-style ARPG, and I want a local AI agent to write most of it. Loot tables, skill trees, procedural dungeons. Step one of many: the dungeon generator. Setup is Qwen 3.8 27B on a single RTX 5060 Ti 16GB, running fully offline. I gave the agent a test file. Solid outer wall. Every floor tile reachable from every other. It passed on the first run. The map was one giant empty room. 90% open floor. No walls means everything is connected. My tests never said the dungeon needed rooms. So I added three more: 20-55% open, narrow corridors must exist, at least 3 separate rooms. Deleted its code. Ran it again. Real dungeon. First try. 36% open, 12 rooms. Next up: monsters, loot, combat.

2 1 30 96K 0

View Details

Lo Merlins @LoMerlins

2 months ago

@0xCapexOG yeah, I hope there is demand at the end of this buildouts cause someone has to pay the bill

1 0 1 15 0

View Details

Lo Merlins @LoMerlins

2 months ago

Full generation video (prompt → code) 👇

0 0 0 66 0

View Details

Lo Merlins @LoMerlins

3 months ago

@0xCapexOG @sydneyrunkle Appreciate it! Yeah, real tests are the only ones that matter to me. Will record and post how it goes.

0 0 1 28 0

View Details

Lo Merlins @LoMerlins

3 months ago

@sudoingX 3.6 27B already felt incredibly solid, even having to squeeze it onto 16GB with offloading. Getting a 3.8 model at that same size is gonna be really interesting to see

0 0 5 882 1

View Details

Lo Merlins @LoMerlins

3 months ago

@0xCapexOG Rising rental prices just make owning even a 16GB card look smarter every month.

1 0 2 35 0

View Details

Lo Merlins @LoMerlins

3 months ago

@0x1mpulse Agree. Start with one useful local agent that actually finishes the job. Complexity comes later.

1 0 2 40 0

View Details

Lo Merlins @LoMerlins

3 months ago

@MiaAI_lab Just tested Qwen3.6 27B on a 16GB card (partial offload) for a coding task — fully playable Snake game on first try, 7/7 requirements. Generation was slower (~9.5 t/s) but the quality jump from 8B/14B was clear. 16GB setups are still very usable in 2026.

0 0 1 142 0

View Details

Lo Merlins @LoMerlins

3 months ago

Even on a 16GB card, the 27b wins. Asked qwen3.6:27b the same snake game checklist (partially offloaded: 20% CPU / 80% GPU). 2021 tokens in 3m37s (~9.5 t/s) — much slower than the 14b’s 40.5 t/s. But the result is clearly the best of the three: 1) Fully playable on first try 2) Correct controls, score, collisions, speed increase every 5 points 3) Game over + restart all working 4) Clean modern look (gradient snake, glowing food, subtle grid) 7/7 requirements hit. 8b → 4/7 14b → 3/7 27b → 7/7 You lose time on generation, but you gain a finished game.

Lo Merlins @LoMerlins

3 months ago

1 0 4 10K 0

0 0 3 9K 0

View Details

Lo Merlins @LoMerlins

3 months ago

Full generation video (prompt → code) 👇

0 0 1 59 0

View Details

Lo Merlins @LoMerlins

3 months ago

Worse than the qwen3:8b🤯 Only 3/7 requirements work. Asked qwen3:14b the same snake game checklist. 1859 tokens in 51s (~40.5 t/s). Much cleaner and shorter than the 8b’s 7442 tokens. Keydown listener is actually present this time. Score display exists. CSS looks fine. Opened it… game over screen appears instantly. Snake never moves (video below). It starts with direction {0,0}, so the first update creates a self-collision on the head and dies before you can press anything. Score also increments by 10 instead of 1 (so the “every 5 points” speed-up is broken too). Bigger model, fewer working features. 27b next. x.com/LoMerlins/stat…

Lo Merlins @LoMerlins

3 months ago

asked qwen3:8b (16gb card) for a snake game against a 7-item checklist. 7442 tokens back. clean code, real css, looks finished. opened it, hit an arrow key. snake ignored it and drove straight into the wall. there's no keydown listener in the final output at all. score display