The difference between "medium" and "xhigh" reasoning effort for Qwen3.8-27B is actually insane.
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
这条记录涉及生成能力或端侧推理进展，适合跟踪模型效率、部署门槛和应用机会。
一位用户正在使用 Unsloth 的 UD-Q4_K_XL 和重新构建的 llama.cpp 在 RTX 2080TI 上测试 Qwen3.8-27B 模型。他们通过 q8_0 量化和 100k 上下文，并使用 MTP 和 --spec-draft-n-max 4 参数，实现了大约 40tk/s 的速度，这略低于 Qwen3.6-27B。…
I'm currently testing out Qwen3.8-27B using Unsloth's UD-Q4_K_XL running a freshly rebuilt llama.cpp. I have a 22GB RTX 2080TI on which I'm able to fit 100k context with q8_0 quantization, and using MTP with --spec-draft-n-max 4 I get about 40tk/s which is slightly less than Qwen3.6-27B but usable enough.
I've been trying to test out some admittedly silly one shot prompts using the llama.cpp webui by asking the model to create fully functional HTML clones of flappy bird, pacman and such, and the difference that changing reasoning_effort makes has been surprising to say the least.
Setting it to "medium" seems to result in barely any thinking at all, a couple thousand tokens max and even less than 3.6-27B. Whereas when using "xhigh seems" I get 15k to 20k thinking tokens at the very least with the pacman example actually hitting 40 thousand fucking tokens.
I'm well aware I can limit the reasoning budget in llama.cpp but I'm wondering if this is expected model behavior or if something is broken somewhere. Any of you guys seeing this?