用了下 DeepSeek-V4-Pro，分享下我对国内模型的看法

好几周前就听说了v4要出，我用过文心、千问、minimax和glm等，比较好用的也就glm5了，平常还是gpt和cc用的多。我同事是坚定的国产模型拥护者，上班一直给我吹v4有多么牛，我先前用过ds v3，吐token是挺快，但是幻觉率挺高。

昨天充了5块，改了下配置文件直接用deepseek-v4-pro了。

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.deepseek.com/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "sk-xxxxxxxxxxxxxxx",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "deepseek-v4-pro[1m]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "deepseek-v4-pro[1m]",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "deepseek-v4-flash[1m]",
    "CLAUDE_CODE_EFFORT_LEVEL": "max",
    "API_TIMEOUT_MS": "3000000",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
    "CLAUDE_CODE_ATTRIBUTION_HEADER": "false"
  },
  "model": "opus[1m]",
  "skipDangerousModePermissionPrompt": true
}

写了个时钟网页，对比下看起来还不错：

有点不敢相信是 DeepSeek 写的，怀疑是 Claude Prompt 发力了，然后整了个 Python 脚本直接 Raw 输出

import json, time, requests

API_KEY = "sk-xxxxxxxxx"
BASE_URL = "https://api.deepseek.com/anthropic"
MODEL = "deepseek-v4-pro[1m]"

PROMPT = "写个在线时钟网页，要求有多功能，使用 Material Design，不准联网搜索。把所有代码写在一个 HTML 文件里，直接可用。"

print(f"正在调用 DeepSeek Anthropic API ...")
print(f"  Endpoint: {BASE_URL}/v1/messages")
print(f"  Model: {MODEL}")

start = time.time()

resp = requests.post(
    f"{BASE_URL}/v1/messages",
    headers={
        "x-api-key": API_KEY,
        "anthropic-version": "2023-06-01",
        "Content-Type": "application/json",
    },
    json={
        "model": MODEL,
        "max_tokens": 32768,
        "messages": [
            {"role": "user", "content": PROMPT}
        ],
    },
    timeout=180,
)

elapsed = time.time() - start
print(f"HTTP {resp.status_code}  |  耗时 {elapsed:.1f}s")

# 保存完整原始响应
with open("raw_response.json", "w", encoding="utf-8") as f:
    f.write(resp.text)
print(f"原始响应已保存到 raw_response.json ({len(resp.text)} 字符)")

if resp.status_code == 200:
    data = resp.json()
    # 找到 text 类型的 content block（跳过 thinking block）
    text_blocks = [b for b in data["content"] if b["type"] == "text"]
    content = text_blocks[0]["text"] if text_blocks else "[无文本输出]"
    usage = data.get("usage", {})
    print(f"Tokens: input={usage.get('input_tokens')}, output={usage.get('output_tokens')}")
    print(f"Stop reason: {data.get('stop_reason')}")

    # 提取 HTML
    html = content
    if "```html" in content:
        html = content.split("```html")[1].split("```")[0].strip()
    elif "```" in content:
        parts = content.split("```")
        if len(parts) >= 2:
            html = parts[1].strip()

    with open("raw_output.html", "w", encoding="utf-8") as f:
        f.write(html)
    print(f"HTML 已保存到 raw_output.html ({len(html)} 字符)")
else:
    print(f"错误: {resp.text[:500]}")

byd 怎么写的更好了

14 个帖子 - 6 位参与者

阅读完整话题