Just wanted to share that I have QWEN 3.6 27B fully working with OpenCode via LMStudio via Unity MCP and it is faster than Coplay / Claude Code / Codex and 100% FREE running on my 4090. Also, seems to be just as good as I gave it a number of fairly complex prompts involving 2-10 decent size scripts across 2-10 scenes and it completed so fast each time I thought it hallucinated but sure enough it was complete. All one shots. I have never experienced so many one shots. It is really impressive and way beyond my expectations.
In terms of speed, this might be more my location as I am in Vietnam right now so agentic AI has to do a lot of round trips to the USA data canters. This might be a viable option for those far away from the AI data centers like me.
It really seems like Open Source is finally a viable option compared to the frontier models and anyone looking to reduce their token costs and/or speed up their sessions with the right hardware really should give it a try. Setup was about a day mostly me learning OpenCode, LMStudio and Unity MCP which is a messy setup to say the least. Hope the next time I set it up I can use an OS level AI Agent as it is mindless and I would rather be working on my VR Theme Park where the real value is.
I’d gladly take another try with yet another local model, though I’d have to look for a quantized version since I only have 16 GB VRAM. My experience, not matter what I tried (and generating only code) was that 50% chance the model just wouldn’t output anything meaningful (errors) or it would take 10x longer than any cloud solution.
40-50 tokens/second at q4 on a 4090 and if I format my prompt like a design document, I get near 100% one shots abd dang fast too. I am truly shocked as I was not expecting it to be this good.
For anyone here looking at Unity AI/MCP-style workflows but worried about subscription/API limits, you might also want to check out Locus for Unity . We recently open-sourced it.
It’s a Unity Dev Agent that supports C# code generation, scene and asset editing, runtime debugging, design documentation, and visual version control for Unity asset changes. The goal is to make AI-assisted Unity development more open and flexible, especially if you prefer using your own setup/API key instead of being locked into one paid tool.
I know this is an insanely good model, especially for a local one, but man, the hardware requirements are crazy. I really wish I could run it, but I just can’t.
