Gemma 4 template request
First of all, sincere thanks for the work invested into Q templates.
Gemma 4 template (even the latest) has been flaky in the vLLM/Opencode even with full models. I hope you will consider putting out your version of it with preserve_thinking support.
Thanks in any case.
Thank you for the confidence. This takes lots of time though, and I am not sure I will be able to. I already put aside other projects for too long to help the community with Qwen chat templates and MTP. I do not even use those models locally... I only did it for the challenge and out of interest, and because I saw Qwen 3.6 27B as a real landmark, that needed several major improvements (speed with MTP, and caht template correctness) to make it properly usable.
Last time I conducted my gemma vs qwen comparisons and benchmarks, qwen won. But that was just after gemma 4 release, and I believe there were initial problems which are now sorted. If I find time, and nobody else does is, I might have a look at the gemma templates, and gemma 4 31b optimisations.
Thank your for a prompt and thoughtful reply. That’s entirely understandable and appreciated.
And - yes - this would be for gemma-4-31b-it which is more of a complement than a competitor to qwen3.6-27b in agentic coding, mainly because of the cons of the sliding window attention of the former: it reasons better than it codes. But it seems like it could be better if tool calling is better.
If you don’t mind me asking, which LLM(s) do you use day-to-day?
That was my conclusion as well when I evaluated both Qwen 3.6 27b and Gemma 4 31b-it:
- Qwen is better for coding and agentic tasks
- Gemma is better for general knowledge, reasoning and multilingual work
I use:
- For coding: GLM 5.1 (I use it heavily, on their Max plan)
- Simple code reviews and general queries: Gemini 3.1 Pro within the Google AI Pro free access
- Image generation: Nano Banana Pro within the Google AI Pro free access