Qwen 3.5 4B - built for KiCAD!!

[mixelpixx]

I have been busy with trying to get 'KiCAD-MCP-Server' to work with locally hosted LLMs. A lot, ok all of them, fail miserably. They cant handle the tool calls, or when they do and it fails then they really cant handle that. So I decided to tackle training an LLM to do what was needed. Through trial an error, and about 2 weeks of training runs, I have a smaller LLM trained on:

KiCAD 10
KiCAD MCP
Model Context Protocol
Tool calling
15K+ long form + thinking + tool calling examples
10K+ edge cases
PCB Design guidance

That isnt all of it. I think I started with 50K+ examples -- I now have a whopping 10M example training corpus, just targeting EE and PCB design via KiCAD (not to mention thinking+tool use). I didnt just 'Fine-tune', I performed 50hrs or more or 'continued pre-training' in order to add the actual domain knowledge.

I have done some testing with the model, and for my tests it passes with 100% in every category. I will have the link up soon, i still have to upload to HF, and get some basic instructions up. In my testing scenario I used llama-serve and just wired it up that way, Llama.cpp runs Qwen 3.5 models, I believe LM Studio does as well. I will put up the GGUF soon, and you can test it via your favorite local hosting method.

I did this because I know that using Claude Code or Desktop is probably costly, and some people may have concerns about proprietary design work and feeding it to one of the big providers. This is ONLY a test. Ideally I would train a larger 32B model which would perform way better than a 4B. And I would probably train different sizes in between.

Ill post a link shortly.

[mixelpixx]

Here is the Huggingface link:

https://huggingface.co/empulse/KiCAD-MCP-Qwen3.5-4B-GGUF

[Kletternaut]

Hey mixelpixx,

sorry for being so quiet lately — I've completely used up my quota and am waiting for the reset.

But down to business:

I tested your LLM empulse / KiCAD-MCP-Qwen3.5-4B-GGUF locally on Win11 / LM Studio with an i9 / RTX4000 workstation (so not exactly a powerhouse setup). The tool connection worked after I configured the MCP. However, the only tool call that actually went through successfully was the project creation. It somehow got tripped up on the rounded corners of the PCB.

I'm attaching the chat history for your reference.

The initial prompt was:
"Create a PCB with 30 x 30 mm and rounded corners, r = 5 mm. Add two LEDs, a power supply, and a button that activates the LEDs. Create a new KiCAD project for this under C:\Users\Tom\Documents\CamStation\AA_PCB\aaa_qwen3.5-4bLLM"

Here is the key error from the debug log:

Model failed to generate a tool call
Failed to parse arguments for tool "add_board_outline":
  params.params requires property "x"
  params.params requires property "y"
  params.params requires property "unit"

Generated tool call attempt:
<tool_call>
  <function=add_board_outline>
    <parameter=shape>rounded_rectangle</parameter>
    <parameter=params>{"width": 30, "height": 30, "cornerRadius": 5}</parameter>
  </function>
</tool_call>

The model correctly identified the right tool and shape type, but didn't include the required fields "x", "y", and "unit" in the params object. It seems like either the tool schema description doesn't make those fields obvious enough, or the model (at 4B params) doesn't have enough capacity to reliably fill all required arguments at once.

The full chat log JSON is available if needed — just let me know, I can send the relevant sections.
Hope that helps with debugging — great project overall, looking forward to seeing it evolve!

Cheers,
Tom

Chat - 2026-03-30 13.29.md

[mixelpixx]

Exactly what i need. Failed calls are gold for training. Its 4B so im not expecting a lot. I have a handful of 3090s and a threadripper, i might try a larger model run. Hopefully more people will share failures. And if people want to share logs of successful tool usage.

This will be a process.

If anyone has GPUs they can grant access to, that would be awesome. Or if you have $12K you found in a jacket pocket.. Im here for you.