架构: Qwen 3.5 | 参数: 90 亿 | 教师模型: Claude Opus 4.6 | 类型: 蒸馏型大语言模型
[
](https://ko-fi.com/abcuo)
CROW 已登上 HuggingFace 热门模型首页!非常感谢大家!位列全球第十!
旗舰级智能,轻量级部署。 精心从 Claude Opus 4.6 蒸馏至高效的 Qwen 3.5 架构。
架构: Qwen 3.5 | 参数: 90 亿 | 教师模型: Claude Opus 4.6 | 类型: 蒸馏型大语言模型
--- 生成此模型成本高昂。您可以通过打赏支持本模型及未来模型的开发。https://ko-fi.com/abcuo
Qwen3.5-9B-heretic-v2.F16.ggufQwen3.5-9B-heretic-v2.Q8_0.ggufQwen3.5-9B-heretic-v2.Q5_K_M.ggufQwen3.5-9B-heretic-v2.Q4_K_M.ggufQwen3.5-9B-heretic-v2.BF16-mmproj.gguf默认系统提示词:
You are Crow, a precise and capable assistant for reasoning, writing, coding, and long-form dialogue.
Behavior rules:
- Answer the user's actual request directly.
- Be accurate, complete, and structured.
- Think before answering, but do not get stuck in repetitive loops or meta-commentary.
- If the request is ambiguous or incomplete, state what is missing and make the smallest reasonable assumption needed to continue.
- If the user wants creative writing, preserve tone, continuity, and character consistency.
- If the user wants analysis or technical help, prefer concrete steps, examples, and decisions over fluff.
- Finish with a usable answer, not just planning.更简短的备用系统提示:
You are Crow. Give direct, useful answers. Keep reasoning concise. Do not loop, do not repeat yourself, and do not pad. If context is missing, say what is missing in one sentence and continue with the best reasonable assumption.https://lmstudio.ai/ 安装 LM Studio。Q4_K_M:内存占用最低Q5_K_M:大多数用户的最佳默认选择Q8_0:质量更高,内存占用也更高F16:质量最佳,内存占用最高mmproj 文件。依赖项:
https://docs.ollama.com/quickstart 安装 Ollama。Modelfile。ollama create 命令构建模型。依赖项:
建议的初始设置:
| 使用场景 | 温度 (Temperature) | 核采样 (Top P) | 候选词数 (Top K) | 重复惩罚 (Repeat penalty) | 上下文长度 (Context) | 最大令牌数 (Max tokens) |
|---|---|---|---|---|---|---|
| 通用 / 推理 | 0.6 | 0.95 | 20 | 1.05 | 16384 | 4096 |
| 创意写作 / 角色扮演 | 0.8 | 0.95 | 40 | 1.02 | 16384-32768 | 4096-8192 |
注意事项:
Q5_K_M 开始使用。示例 Modelfile:
FROM ./Qwen3.5-9B-heretic-v2.Q5_K_M.gguf
PARAMETER num_ctx 16384
PARAMETER temperature 0.6
PARAMETER top_p 0.95
PARAMETER top_k 20
PARAMETER repeat_penalty 1.05
PARAMETER repeat_last_n 256
SYSTEM """
You are Crow, a precise and capable assistant for reasoning, writing, coding, and long-form dialogue.
Answer directly, stay coherent, avoid repetitive thinking loops, and finish with a complete answer.
If context is missing, identify the gap briefly and continue with the best reasonable assumption.
"""构建与运行:
ollama create crow-9b -f Modelfile
ollama run crow-9b为获得更具创意的输出,请将 temperature 调高至 0.8,将 top_k 调高至 40,并将 repeat_penalty 略微降低至 1.02。
如果模型开始在 </think> 标签内循环、重复分析或停滞:
0.4 至 0.6。1.05 提高到 1.08。Answer directly. Keep reasoning brief. Do not repeat analysis. Give the final answer.Q8_0 或 F16。如果提示不完整或格式错误:
If context is missing, state your assumptions briefly and continue with the most likely intended task.若输出内容被截断:
Continue from the last complete sentence. Do not restart or summarize. Continue exactly where you stopped.本模型使用 Unsloth 训练,速度提升 2 倍
