【RTX Pro 4500 Blackwell】Qwen3.8 27B NVFP4 MTP在dsh的真实表现
前言
之前已经在LM Studio运行了Qwn3.8 27b，并且跑成功了图像识别：
【RTX Pro 4500 Blackwell】Qwen3.8 27B NVFP4 MTP在LM Studio的性能表现
https://lizhiyong.blog.csdn.net/article/details/163834176
【RTX Pro 4500】Qwen3.8 27B NVFP4在LM Studio挂载多模态模型
https://lizhiyong.blog.csdn.net/article/details/163834242
最近流行使用 deepseek harness 做AI Agent，在内网开发效果还不错，正好试下qwen3.8 27b的效果
运行模型
加载模型
依旧是懒人做法，在LM Studio启动：
虽然RTX PRO4500 Blackwell仅有32GB显存，得益于NVFP4及Q4量化KV cache，可以看出能够拉满262k的上下文。虽说比deepseek v4 pro的1M上下文弱太多了，但是已经初步具备了真正意义的生产力。
验证模型可用
java
用户让我看一张图片，没有具体提问，所以我应该主动描述并解读这张图。我需要仔细观察细节。
观察图像：一只豹子（leopard/cheetah-like, spotted big cat）戴着墨镜，躺在 a beach on a green towel/mat. It's holding a newspaper with one paw, sipping a red fruit drink (looks like watermelon juice with orange slices, lime, mint, ice) through a straw. Background: tropical beach, palm trees, turquoise sea, blurred people sunbathing/swimming, an umbrella.
The green mat has Chinese text on it: "虎鲸不是鱼" which translates to "Orca is not a fish" / "Killer whale is not a fish". This is a known phrase/meme in Chinese internet culture? Let me think. "虎鲸不是鱼" --- there's a popular saying or maybe a book title? Actually I recall "虎鲸不是鱼" might be related to a meme or just a factual statement used humorously. Hmm, but why would it be on the towel? It's likely an AI-generated image with some random text inserted, possibly as a watermark or easter egg, or the user wants me to notice the text.
Wait, maybe this is a test of whether I can read the Chinese text and interpret the humor/oddity. The image is clearly AI-generated (the leopard's anatomy, the way it holds the newspaper, the drink details). The text "虎鲸不是鱼" is somewhat incongruous with the scene---a leopard on a beach. So there's an absurd juxtaposition.
Let me think about whether "虎鲸不是鱼" has special meaning. In Chinese, 虎鲸 = orca/killer whale. It's a mammal, not a fish. The phrase might be a reference to something. Actually I think there's a viral thing: people say "XX不是YY" as a humorous correction. Or it could be the title of a children's book / science fact card. Hmm.
Actually, I recall now: There's a Chinese internet meme or a brand? Let me search memory... "虎鲸不是鱼" might be associated with a specific account or slogan. Not sure. I shouldn't over-claim. I'll mention it as the text on the towel and note its literal meaning, and that it adds to the surreal/humorous tone, possibly an easter egg.
The user just said "帮我看下这个图片" = "help me look at this image". So they want a description/analysis. I should provide a clear, friendly breakdown: what's in the image, the AI-generated nature, the humorous concept (a leopard living its best life on vacation), and point out the Chinese text and its meaning, plus any oddities/inconsistencies that reveal it's AI art.
I'll structure: overall scene, the cat details, the drink & newspaper, background, the text on the mat, and a note on style/AI tells + the comedic vibe. Keep it conversational, not too long, maybe with a couple of bullet points or short paragraphs. Use Chinese since user wrote in Chinese.
Let me look at the cat species: rosette spots = leopard (or jaguar). The face looks more like a leopard/cheetah hybrid but rosettes suggest leopard/jaguar. I'll say 豹子（花豹/美洲虎一类）. Fine.
Tells of AI: the newspaper text is gibberish, the straw goes into mouth oddly, the paw holding paper has weird finger structure, the drink fruit slices look slightly off, background people are blurry blobs (intentional depth of field though). The Chinese characters on the mat are actually rendered fairly correctly which is notable---AI often messes up CJK. "虎鲸不是鱼" looks correct: 虎 鲸 不 是 鱼. Yes all five chars look right. Good, I'll note that the text is legible and correct, which is a nice touch or intentional insertion.
Tone: light, observant. End with maybe a question inviting what they want (e.g., do they want me to edit it, explain the meme, etc.)? The instruction says response must stand alone. I can offer follow-up options briefly.
Let me write in Chinese.思考后输出：
java
这张图挺有意思的，是一张AI 生成的"豹子度假"概念图，整体走的是反差萌 / 拟人化搞笑路线。我帮你拆一下：
画面主体
一只花豹（斑点是大猫里的 rosette 花纹，偏花豹/美洲虎那一类）戴着墨镜，懒洋洋地趴在沙滩上，后腿翘着、尾巴甩在一边，一副"老钱风度假"的松弛感。
它一只手（爪子）举着一份报纸在"看"，嘴里叼着吸管，正在喝一杯红色水果冰饮------杯里能看出西瓜块、橙子片、青柠和薄荷叶，配冰块，很像一杯 watermelon cooler / 莫吉托风格的饮料。
背景
典型热带海滩：椰子树、 turquoise 的海水、远处的山影，还有虚化处理的游客在游泳、晒日光浴，右边一把红黄遮阳伞。景深做得不错，把焦点全压在豹子身上。
最关键的彩蛋------垫子上的字
绿色沙滩巾上印着一行中文：「虎鲸不是鱼」（意思是"虎鲸/逆戟鲸并不是鱼，而是哺乳动物"）。
这句话和画面里的豹子其实没有逻辑关系，属于故意塞进去的"梗/水印式"文字，制造一种荒诞感。值得注意的是这几个汉字字形基本都生成对了，这在 AI 图里算难得（AI 经常把中文写崩），所以更像是后期特意加上去或专门调过的。
能看出是 AI 的小破绽
报纸上的"正文"是一团糊掉的伪文字，没有真实可读内容；
叼吸管的那段嘴部结构、以及举报纸的爪子指节，仔细看有点不太对劲；
杯子里水果切片的边缘和光影略"塑料感"。
整体观感：一只把日子过成精的豹子，在沙滩上边喝果饮边看报，配上那句没头没尾的"虎鲸不是鱼"，幽默又 surreal。
你是想让我解释这句"虎鲸不是鱼"的梗/出处、判断它是不是 AI 做的，还是想基于这张图改点什么 / 写配文？告诉我方向，我接着帮你弄。速度还不错：
思考了24s，token速度37个/s。
集成到DSH
懒人做法，使用Deepseek Harness Desktop版本，省的安装各种前端依赖包了。。。在内网已经证明了这种版本也可以左脚踩右脚，实现赛博飞升。
开启Server
需要在LM Studio开启服务才能使用
配置模型提供方
需要在Deepseek Harness Desktop配置本地模型为自定义的模型提供方：
唯一有用的就是：http://127.0.0.1:1234/v1，其余的随便填点什么即可。。。
验证效果
新建个工作空间，选用本地模型测试。
然后它表示：
java
I found one image in D:\test\1:
Qwen-Image-2512_00086_.png
Format: PNG, ~3.3 MB
Dimensions: 1920 × 1088 (landscape)
However, I can't actually see it --- the current model (hauhaucs/qwen3.8-27b-hauhaucs) doesn't support image input, so read_image was rejected by the harness. All I could do is read its file metadata.
To describe what's in the picture, you'd need to switch this session to an image-capable model (e.g., a vision-enabled Qwen or similar) and ask me again --- then I can open it directly and tell you what it shows. Want me to try anything else with it in the meantime?显然模型是被调用了，但是多模态能力依旧欠缺！！！
启动图片识别功能
此时需要按照豆包提供的建议，修改：
C:\Users\zhiyong\.dsh\settings.yaml：
yaml
ui-onboarding:
welcomeNoticeVersion: 2026-08-13.1
llm-pi-ai:
providers:
siliconflow:
displayName: 硅基流动 SiliconFlow
apiKeyEnv: SILICONFLOW_API_KEY
api: openai-completions
baseURL: https://api.siliconflow.cn/v1
models:
- id: meituan-longcat/LongCat-2.0
name: LongCat-2.0
- id: deepseek-ai/DeepSeek-V4-Pro
name: DeepSeek-V4-Pro
- id: deepseek-ai/DeepSeek-V4-Flash
name: DeepSeek-V4-Flash
input: [text, image] # 👈 加上这一行！声明模型支持图片
lmstudio:
displayName: LM Studio Local
apiKeyEnv: LMSTUDIO_API_KEY
api: openai-completions
baseURL: http://127.0.0.1:1234/v1
models:
- id: hauhaucs/qwen3.8-27b-hauhaucs
name: qwen3.8-27b
input: [text, image] # 👈 加上这一行！声明模型支持图片
agent-default-model:
provider: lmstudio
model: hauhaucs/qwen3.8-27b-hauhaucs
dsh-community-market:
sources:
- sourceRecordId: 6ded514f-886e-4952-9bd2-ec30ecdc658c
registrationKind: built-in
adapterId: market.dsh-1024store-v1
providerId: com.deepseek1024.catalog
builtInProviderKey: dsh-1024store
enabled: true
order: 0
dsh-desktop:
mode: extended
macosMaterial: transparent
windowsMaterial: mica
openBrowser: false
networkExposure: loopback
dsh-desktop-notifications:
enabled: true
notifyOnTurnCompletion: true
notifyOnTurnFailure: true
notifyOnJobCompletion: true
notifyOnJobFailure: true
ui-conversation:
busyEnter: steer
ui-theme:
preference: dark由于deepseek v4 flash4.1其实也支持多模态了，所以一起改了！！！这样就能实现图片识别：
此时可以调用模型识别到图片了！！！
目前的效果：
java
2 轮 · 7 步| LLM 51.7秒 · 工具调用 2秒| 首 token 平均 1.8秒 · 44 tok/s| 缓存命中 0%| 输入 66.8K tok · 输出 1.7K tok实际测试编程能力
Qwen3.8 27b小模型号称编程能力已经战胜了上一代deepseek v3 671b，使用一个贪吃蛇游戏试试效果：
在运行了10多分钟后：
还是能玩的。
消耗145k的token。
使用成本
推理时整机功耗360w，公寓电费1块2，综合成本3元/M，比云端的deepseek v4.1 flash还贵，也更不聪明。不过好处是小型任务可以token自由了，也不会被模型拒绝。
cmd
C:\Users\zhiyong>nvidia-smi
Sat Sep 12 00:20:04 2026
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 597.06 Driver Version: 597.06 CUDA Version: 13.2 |
+-----------------------------------------+------------------------+----------------------+
| GPU Name Driver-Model | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA RTX PRO 4500 Blac... WDDM | 00000000:C1:00.0 On | Off |
| 49% 78C P1 199W / 200W | 31203MiB / 32623MiB | 89% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+子任务使用情况
模型可以自主规划subtask。
实际用作的逆向的效果及细节不便展示，逆向笔记系列第一篇：AI 自动化逆向（一）最贵的错误是「看起来像结论的错答案」
转载请注明出处：https://lizhiyong.blog.csdn.net/article/details/167813953