Running wsl2, sglang with gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090, medium and dspark, 128k context, with opencode, about 150 tok/s. It does seem faster than unsloth, unsloth didnt support dspark.

Barely fit when windows is stealing 2gig vram from wsl. But very very usable.

Switching from xhigh to medium roughly halved the time it took on a task for me. It's still a very verbose and long thinker, even so. It's hard to say for sure what the impact on output quality is at medium, but it seemed pretty close.

Because it's so nice to see page after page after page of... but what if... let's consider...

If you use a gguf, you can extract the template from the Qwen3.827B model using the following script https://github.com/alainnothere/llama.cpp/blob/disk-cache-ev..., use it like

ggufGetTemplate.py MODEL.gguf [OUT.jinja]

modify the template to your liking, and then use https://github.com/alainnothere/llama.cpp/blob/disk-cache-ev... to push it into the gguf.

ggufSetTemplate.py MODEL.gguf TEMPLATE.jinja [OUT.gguf]

This one https://github.com/alainnothere/llama.cpp/blob/disk-cache-ev... has the default effort set to medium, so you can just push it into your gguf and use it.

Interesting things, the "effort" is an instruction telling the model how to behave, medium doesn't have anything (at least on the version of q3827 that I have) so you can set whatever you like.

These are the defaults

low: "Reasoning effort is set to low. Keep your thinking brief and focused, moving directly to the conclusion without unnecessary elaboration."

xhigh: "Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer."

medium: nothing is set here... in the template above I set the instruction to: "Reasoning effort is set to medium. Think through the main considerations and verify the key steps, but do not be exhaustive: once the essentials check out, commit to the answer." and I like the resulting behavior.

If you wondered what happens if you tell the model to answer like a commedian, stop wondering and start experiment.

Here are the results of togling the effort per message (you have to use my version of llama.cpp in the link to try that) you can see it here: https://github.com/alainnothere/llama.cpp/blob/disk-cache-ev...

Happy hacking.

I just end up setting llamacpp --reasoning-budget so it doesn't really matter how much it wants to think, it's forced to return and do whatever fruitful thing it can do with the budget it has. I tie the --reasoning-message into tool use or subagents or dynamic context pruning, so there's almost always something it can do to unblock it's thinking