What would you use for local coding assist on a "weak" machine (6GB VRAM 32 GB RAM) - light FE coding, no architecture. is QWEN3 good enough?

Posted by vishnoo@reddit | LocalLLaMA | View on Reddit | 4 comments

so as it says, I am not a FE eng, but want to do some light FE work
I don't need the smartest model but need to get some work done.
I ran out of tokens (20$ a month) for the week on day 2, so thinking of running something local
I tried serving QWEN3 with ollama and connecting codex to it, but it was clunky at best.

I figured I'd ask the experts

so local windows machine, I ran it on WSL, but codex then had issues accessing the local directories. is it better to run it in PowerShell (shudder)

gemma4:26 (quantized) also sort of fits but provided worse results.

to sum up
1. WSL vs windows native
2. codex? (claude-code blocked local models) opencode?
3. qwen? gemma?

[-]

mr_Owner@reddit

Check byteshape quants, you'll be surprised by their 9b and 30b coder variants.

EffectiveCeilingFan@reddit

If you're using enough tokens that you can run out for the week on the second day, then I'm sorry to say that your machine will feel unusably slow. The best model would probably be Qwen3.5 35B-A3B if you want to try something, but it's going to take potentially minutes before it even begins thinking, let alone actually responding. Don't use Gemma4 yet, fixes are still coming daily, it needs another few weeks before it's ready for a serious test.

Consider looking into a more lenient API plan like Kimi Code, GLM Coding Plan, or the MiniMax Token Plan. You'll get much more usage for your dollar.

vishnoo@reddit (OP)

between the 3 options, which do you like best?

hmm, so splurge on max.... yeah, i guess.