Qwen 3.8 Locally: How Much Can You Actually Trust It?

Qwen 3.8 27B is a new open model, and instead of asking whether it can replace Opus or Gemini, I wanted to answer a more practical question: how complex can a real task be before a local model of this size stops being reliable?

🧩 MODEL
Qwen 3.8 27B, Q4_K_M quantization

————————————–

LINKS
📬 My newsletter (Zero to MVP Weekly) — https://weekly.blokhin.us
📋 Prompt collection used for the tests — https://github.com/w512/Prompt-Vault
🛠 Test 3 app (Feed Aggregator) — https://github.com/w512/Feed-Aggregator
❤️ Patreon — https://patreon.com/ZerotoMVP

————————————–

⏱️ TIMESTAMPS:
00:00 What I’m actually testing
01:01 Setup: hardware, Ollama, and the Pi agent
02:33 Test 1 — clock, stopwatch, and timer
04:06 Test 2 — Pomodoro timer, no plan
05:44 Test 3 — full-stack news aggregator in Nim
06:47 Where the model hit its limit
08:15 Running the finished prototype
09:45 Qwen 3.6 vs Qwen 3.8, and what I’d trust it with

————————–…