The model
Qwen3.5 4B
A small open-source model from Alibaba's Qwen team, released March 2026. It's small enough to run on a laptop processor, with no graphics card.
- LicenceApache 2.0 (free and open)
- Good atExplaining, summarising, tidying up text
- Backup modelIBM Granite 4.2 3B (faster, more literal)
- Runs withOllama + Open WebUI
Model specs
By the numbers
Measured on my server, not taken from a brochure.
- Size
- ~4 billion parameters
- Download
- 3.3 GB
- Memory in use
- 3.4 GB
- Memory per chat
- 8,192 tokens (~6,000 words)
- Writing speed
- ~8 tokens/s (~6 words/s)
- Reading speed
- ~28 tokens/s
- Thinking mode
- Off, so it answers right away
Try it
Ask my AI
AI
Hi! I'm a small AI running on a laptop in Faiyaz's homelab. Ask me about his projects, or anything else. I'm slower than ChatGPT, so give me a few seconds.
It answers one person at a time, so replies can take 15–30 seconds. It can make mistakes, so don't trust it for facts that matter, and don't share personal information.
The hardware
Server specs
A Debian container on my Proxmox homelab.
- Host machine
- HP EliteBook 840 G5
- Processor
- Intel Core i5-8265U
- Cores for the AI
- 4
- Memory for the AI
- 7 GB
- Graphics card
- None, CPU only
- Operating system
- Debian 13
- Hypervisor
- Proxmox VE 9.2
Guestbook
Comments
Tried the AI? Tell me how it did.
Past comments