I tried some local coding with qwen 3 and spent hours “arguing” with it and did not get very far. I am curious if anyone had some experience here that they wanted to share.
I have also used bolt.new, Lovable, and most recently, before qwen 3 coder, I used Gemini CLI to fix some SEO stuff on a website. It sometimes feels like TRUE magic and you can fly. Other times it feels like I am talking to the most epically retarded intern the universe ever made. Like truly can’t wipe its own ass correctly. Sorry if that is crude, but it can be frustrating using these tools.
Just stop anthropomorphizing the tool. Better responses require better prompts so instead of arguing with the tool ask yourself how you can change your prompts to better provoke desired responses.
It’s also often helpful to bail and start a new context. Once it’s been polluted with nonsense it can be difficult to move forward.
1 Like
I switched to using qwen3:30b after it came out on ollama last week. I find it is pretty good. I’ve been using it with Open WebUI and Cline.
I have been using @ubergarm Qwen3-Coder-30B for a few days and been finding it great. When it struggles, I then jump over to Claude. I have found my Claude usage has dropped way down and I prefer the shorter answers from Qwen3-Coder rather than the verbosity of Claude
Although, it seems quality is entirely subjective, I am not looking to have my llm one shot features, I am using to fault find and learn. As @tk11 said, don’t argue with it if you are not getting the answers to you want, it isn’t going to learn a lesson. Just start a new session and change your prompt.
https://huggingface.co/ubergarm/Qwen3-Coder-30B-A3B-Instruct-GGUF
I am very curious to try the full 480B model to see if it is any better, but I would need to buy more RAM and make my machine single purpose, so I will give it a bit more time to better evaluate the 30B model.
1 Like
While a bit larger than the Qwen3-Coder-30B-A3B quants, I just uploaded some GLM-4.5-Air quants here: ubergarm/GLM-4.5-Air-GGUF · Hugging Face which are a good “middle sized” model.
Also I heard about a new Qwen3-Coder that was distilled from the larger 480B supposedly with better capabilities (according to random reddit person hah): https://www.reddit.com/r/LocalLLaMA/comments/1mizz4c/i_distilled_qwen3coder480b_into/
Glad you can generate more tokens locally and ease off the anthropic sauce! 
2 Likes
I was able to get it to work in Roo Code (agent mode, sort of). It is surprisingly difficult to use it as an agent, I tried with Copilot - not supported, while continue.dev has issues with tool calling (only works through certain clients like LMStudio)
Yes, this is one of the other models I was eyeing up. I don’t have an intuitive sense how model performance scales with size. Obviously the larger the better, bit is twice the size twice as good? Then mix in the different quants and all the rest! 
On the link that you posted, someone asked a question that I have been wondering too.
The other thing that gets weighed up in my assessment of usefulness is basic tool calling. Much like @Arkansaw, I have not had huge success with Qwen3-Coder. It seems that this is more difficult to know what is going wrong, there is more than just the model in play here. I am using opencode and I have tried Avante with mixed results. I would assume that tool calling works well with qwen coder but have not tested yet. All part of the fun of learning to self host.
My computer is serving a few duties with proxmox. Nothing that I can’t move elsewhere but ideally would prefer not too. I was thinking about getting one of the 128gb RAM kits linked to on the R1 thread to test the larger models at a slower inference speed just to evaluate and then complete a full install if worth it. Certainly appreciate all the models and info you are sharing!
There has been some progress with tool-calling features on ik_llama.cpp and currently a new PR to support --jinja which could be useful if anyone wants to try it out and report on github: add jinja template support by firecoperana · Pull Request #677 · ikawrakow/ik_llama.cpp · GitHub
1 Like
If you use vLLM you will need to add the following args to your server command: --enable-auto-tool-choice --tool-call-parser qwen3_coder. Then it works really good, using 30b version just fine with Roo Code.
This model (Qwen3-coder) will also double as an agent and handle autocomplete already if hosted. A base model is not provided (Only the fine-tune instruct), but it’s still the FIM version. You will need to configure your autocomplete prompt template: <|fim_prefix|>{{{prefix}}}<|fim_suffix|>{{{suffix}}}<|fim_middle|>. So far this is one of the better FIM models.