It's been a little over 2 months since Qwen3.5 was released? In just 2 months, Qwen3.6 has been released.
It's the Qwen3.6-35B-A3B moe version, and it has higher evaluation scores compared to other models (Qwen3.5, Gemma4).
It has risen to almost the same score as the Qwen3.5 27b dense model that I mainly use for coding.
I tested it by loading it on my rtx3090 24g card, and I tested it by loading a q4 quantized model.
When I did a simple output speed test, the decoding output speed comes close to 100 tok/s,
By extending the context as much as possible to 200k and offloading to cpu layer,
I can achieve decoding speed of up to 70tok/s.
Usually when coding, I use the Qwen3.5 27b dense model with 100k context and it gets the job done,
but Qwen3.6-35B-A3B has overwhelmingly faster output speed... if the coding performance turns out well, it seems like I could switch to it.
Also, from what I saw in reddit conversations, it looks like a Qwen3.6 model corresponding to the Qwen3.5 27b dense model is also in preparation.
Since unsloth is already distributing gguf, if you're doing local llm, please check it out.~
