# a qwen on the home card.

Qwen3.5-35B-A3B on an RTX 3090, running through llama.cpp on Windows and connected to Claude Code.

Author: Vlad Kostoglodov. Language: en.
Archive date: 2026-02-26. Current version prepared: 2026-10-14.
Date basis: Date of the original Telegram post.
Status: ready.
[Profile and technical work JSON](https://vk0st.com/facts.json): public profile, concrete work and source references.
[Human-readable article](https://vk0st.com/writing/qwen-on-the-home-card/).

Qwen3.5-35B-A3B fits on my home GPU and runs surprisingly briskly. I got it running through llama.cpp on an RTX 3090, on Windows. The Windows part earned its own parenthesis in my original post.

![An outlined two-fan GPU inside a computer frame supports a local neural structure and small terminal.](https://vk0st.com/media/articles/qwen-on-the-home-card-v2.webp)

A model on the home card, with the experiment contained in the local machine.

In that February run, I reported roughly **100 tokens per second**. My post didn't include the quantization, context length, or measurement setup, so treat that as an observation from one run rather than a benchmark to compare against.

[This was the GGUF model link](https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF) I shared alongside it.

Then I added it to Claude Code. My update ended with “let's try it,” which is exactly where this little story stops: a model running on the home card, connected to a coding tool, with the useful part still to find out.