Early in the morning, qwen3.5 gguf came out and I tried using it lightly on my local machine.
I have an RTX 3090 with 24GB VRAM and 32GB system RAM configuration.
Recently hot glm 4.7 flash and qwen3 coder next were tested with q4 models while contorting myself to test performance.
qwen3 coder next is slow in speed
but it was good.
If I'm satisfied with roughly 23tok/s output tokens, I was going to focus on coding, but,,,
The q4 quantized model of qwen3.5 35b a3b easily exceeded the record of qwen3 coder next.
Roughly 65+tok/s output tokens without CPU offloading.
The performance is.... good enough to create light web games in one go.
What's amazing is that this is a multimodal model.
I haven't tested images yet, but..
I'm looking forward to when the qwen3.5 coder model comes out.
Chinese AI is amazing.
===========================
Added 26.3.8: I compared the qwen3.5 35b a3b model and the qwen3.5 27b dense model,
and in terms of performance compared to the qwen3.4 35b model, the qwen3.5 27b dense model has about 3 times better scores,
and the speed is about 3 times slower.
Ultimately, I chose the qwen3.5 27b dense model on my system,
and for those with insufficient system performance VRAM, it would be better to use the qwen3.5 35b a3b model.
