MacBook Pro M1 (64GB) + VSCode + Roo + LM Studio + Qwen3.6-35B-A3B-Q6_K.gguf = 😞
Posted by ExplorerWhole5697@reddit | LocalLLaMA | View on Reddit | 4 comments
I've tried the setup in the title today for some vibe coding (ctx=262144, temp=0.6). I must be doing something wrong because it doesn't really work for me.
For example, I have a web based product configurator that uses SVG images extensively, and I told it to hide a specific element that is present in all SVG:s. Super simple. We're already manipulating the SVG:s so I expected it to do something like getElementByID(layerName).style.display = none.
Nope. First it tried to delete the element from the SVG files themselves. Then it wanted to inject a new CSS rule into loaded SVGs to hide the element. Then it tried to inject an inline CSS style using regex...
Of course, these are all "valid" approaches, but not at all what I wanted. I tested some commercial LLM:s and they all nailed this perfectly.
I've also tried Qwen3.6-35B on some more challenging (but still reasonable) problems. For example, I asked it to plan and implement basic undo/redo functionality. Plan looked alright, but now it's been running in circles for an hour trying to implement it.
What can I do to improve things?
- Should I lower my expectations?
- Try another quantisation?
- Change model?
- Change configuration, prompt or software stack?
uti24@reddit
This model tends to loop. Try enabling presence_penalty.
This model is kinda great at one shots, but it falls apart on longer tasks.
"we have model at home" vibe. Yeah, before that we had even worse so this is still considered great for free local model.
ps5cfw@reddit
It Is true that the model really likes to loop on more complex tasks, which Is why I Always tend to steer It toward adding logging to Better assess the situation (for coding scenarios at least)
If provided with enough information and clear requisites, It works well enough that I am using It Daily for professional purposes, but you really gotta give It a lot of information (I am Always using at least 100K worth of information context basically)
ExplorerWhole5697@reddit (OP)
Agreed; qwen are doing miracles, and I can *sense* the greatness of this model. We're so close now.
I will look in to presence penality, thanks.
Awwtifishal@reddit
Use the 27B dense instead of the 35B MoE. Also I had better luck with kilo code than with roo code with local models. Kilo supports native tool calling.