ulofi Forum › Q&A
Qwen 上新 Qwen3.8-Flash-Next-FP8 权重
No answers yet — you could be the first.
Qwen 在 HuggingFace 放出了新模型 Qwen3.8-Flash-Next-FP8,FP8 量化版。页面目前 128 赞、2219 次下载。只放出了 chat template 代码,参数量和跑分一个字没提,想看门道只能扒模板。
1. 模板里内置了三档推理强度,叫 reasoning effort,可选 xhigh、medium、low,默认 xhigh。xhigh 对应的指令文本是
> 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.
low 对应的是
> Reasoning effort is set to low. Keep your thinking brief and focused, moving directly to the conclusion without unnecessary elaboration.
medium 只出现在允许列表里,模板没给它配指令文本。
2. 工具调用写成 XML 格式,写法是
> <tool_call>
> <function=example_function_name>
> <parameter=example_parameter_1>
> value_1
> </parameter>
> </function>
> </tool_call>
模板还写死了规则:函数调用前可以带自然语言说明,调用后不准再补文字。
3. 视觉输入靠占位符,图片是 <|vision_start|><|image_pad|><|vision_end|>,视频是 <|vision_start|><|video_pad|><|vision_end|>。系统消息里塞图片或视频,模板直接报错。
4. enable_thinking 默认开着,关掉后模型吐出空的 <think>\n\n</think>。
原文:https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP81
Log in to replyLog in