2-bit Qwen3.6 Tool-calling is amazing!!
Hey guys, feel free to run and train Qwen3.6 in Unsloth Studio. The tool-calling is quite amazing, even for 2-bit.
GitHub: https://github.com/unslothai/unsloth
Guide: https://unsloth.ai/docs/models/qwen3.6#unsloth-studio-guide
Below, the 2-bit Qwen3.6 GGUF made 30+ tool calls, searched 20 sites and executed Python code.
4-bit example:
thanks i was on the fence about trying the 2 bit i wish unsloth studio was easy to set up like lm studio and had a app or something maybe that would hurt performance too much idk looks great tho
thanks i was on the fence about trying the 2 bit i wish unsloth studio was as easy to set up like lm studio and had a app or something maybe that would hurt performance too much idk looks great tho
We're gonna release the desktop app this month! Needs more testing π
thanks i was on the fence about trying the 2 bit i wish unsloth studio was as easy to set up like lm studio and had a app or something maybe that would hurt performance too much idk looks great tho
We're gonna release the desktop app this month! Needs more testing π
Could you please add inference support for Intel Arc GPUs to Unsloth Studio? π
without any doubt, you make the best gguf quants ( I did not try other forms). I always go to the UD_Q4_K_XL as it is fast with near lossless experience.
Re Qwen3.6-35B: to be honest, I did not see any meaningful difference compared to the 3.5 , not sure what qwen is talking about with all the way higher benchmarking compared to older version.
Re Qwen3.6-35B: to be honest, I did not see any meaningful difference compared to the 3.5 , not sure what qwen is talking about with all the way higher benchmarking compared to older version.
It seems to perform even better with tool calling for me!
thanks i was on the fence about trying the 2 bit i wish unsloth studio was easy to set up like lm studio and had a app or something maybe that would hurt performance too much idk looks great tho
If you use llama.cpp, there's also a webui there with llama-server :)
This model is sort of good even at Q2_K_XL i had it make a website with this prompt (make a mk ultra website in html make it amazing ) and i got 16k tokens 1.7k lines of code I've tested online it seems produces way better websites than gemma 4 and in LM studio at least with the DuckDuckGo and visit website tool it's far less lazy with tool calls it could just be the way they're configured but even when I tested gemma4 in AI studio they hardly ever use the Google grounding and the higher benchmark scores are more than likely because of training for tool use πsorry for rant [{DATE: 4/19/26 }] i recommend to put this in the System Prompt
and one shot space invader game at Q2_K_XL