2-bit Qwen3.6-27B GGUF made 26 tool calls, triaged 15 GitHub issues and fixed, tested + reproed our repo’s 3 latest issues. 🔥 Try this locally in Unsloth Studio with just 12GB RAM. Studio also has a new look! GitHub:

Use Unsloth Studio’s built-in code execution sandbox + structured output (Outlines/JSON mode) for your agent. Forces clean reproduction steps + patch diffs every time. No more hallucinated fixes. Tested on 27B …purity jumps noticeably.

bro 12GB RAM handling 26 tool calls + 15 gitHub issues triaged and fixed is exactly what local inference should be. qwen 3.6 runs solid on cheap VPS for daily stuff too 💯

Is the plan that running inference of Unsloth models through Unsloth Studio will be better than through say llama.cpp?

Wow, that’s nice! How do you make it work in only 12GB? Q4? What else? I thought the RTX 3090 would be the bare minimum.

Someone should host this and give us an endpoint alrey

It’s worth? Better 2bit 27b or 9B q5?

crazy gonna try this out. if this works reliably im gonna fall in love

cc @MatthieuTalbot ça devient sérieux là

Nah at this point this is some black magic shenanigans right here

I'm trying to understand the language here because you mentioned that it does 26 toolcalls, but I don't see a world where any model can triage 15 GitHub issues, fix a few of them, reproduce them with tests first, and verify the fix in 26 toolcalls. That doesn't make any sense.

@UnslothAI will you be releasing an nvfp4 version of qwen 3.6? would love to try it on blackwell

How are you all running this? Is this via your new desktop app?

@nisten Do you think 2-bit 27B or 8 bit 9B better?

absolute magic, thanks

Whoa 😮

12 gb of ram?

Wth with only 2 bits 😲

crazy man will try it asap

the real unlock here is not the 26 tool calls or triaged issues, but the 3 fixed ones that actually matter to users.

Qwen 3.6 is the absolute best hell yeah! ☝️😎

Dgx spark pls 👍

Bruh, is that mean I can use it locally with an rtx 4060 ?

2 bit wtf?

where's the link? I only see 4bit in your HF. Thanks

I just downloaded this exact model from the video. I have 12GB of VRAM and it doesn't go above 4K context. Something is wrong here. Can you explain? @UnslothAI

2 bit? 🫪

UD Q2 KL on 3060 12gb = 6 tps. Local model are great for high request but low tokens usage.

When will the app be available on Mac? :(

Will unsloth studio can be able to serve model with endpoint?

Does unsloth studio install already on the today released Ubuntu 26.04 LTS?

2-bit running actual agentic workflows on 12GB is the real story. a year ago you needed 48GB for something half as capable

How much context? What KV cache quant?

Are you saying I can run this on a Mac Mini ?

@juanorion lo que te comenté de la versión de 2 bits