跳至主要內容

Qwen3.6-27B LiteLLM / Claude Code 配置基準

Backend: vLLM @ http://<VLLM_HOST>:18000/v1 Model: qwen3.6-27b-fp8

Alias 總覽

Alias	用途	thinking	temp	top_p	max_out	max_in	timeout
`qwen3.6-27b-code-act`	Claude Code / OpenCode / Codex 改檔	off	0.5	0.9	16384	245760	600
`qwen3.6-27b-code-think`	規劃、debug、架構分析	on	0.6	0.95	24576	237568	600
`qwen3.6-27b-stable`	OpenClaw、通用分析	on	0.6	0.95	16384	245760	600
`qwen3.6-27b-strict`	JSON / tool / 風控規則	off	0.5	0.9	8192	245760	600
`qwen3.6-27b-fast`	摘要、翻譯、快速問答	off	0.7	0.8	4096	245760	600
`qwen3.6-27b-report`	長篇研究、報告	on	1.0	0.95	32768	229376	900

統一價格: input $0.32 / 1M,output $3.20 / 1M Backend context cap: 262144(所有 alias 滿足 max_in + max_out ≤ 262144)

LiteLLM 配置範例

{
  "model": "vllm/qwen3.6-27b-fp8",
  "api_base": "http://<VLLM_HOST>:18000/v1",
  "custom_llm_provider": "hosted_vllm",
  "max_tokens": 16384,
  "max_input_tokens": 245760,
  "temperature": 0.5,
  "top_p": 0.9,
  "timeout": 600,
  "extra_body": {
    "top_k": 20,
    "chat_template_kwargs": { "enable_thinking": false }
  }
}

{
  "model": "vllm/qwen3.6-27b-fp8",
  "api_base": "http://<VLLM_HOST>:18000/v1",
  "custom_llm_provider": "hosted_vllm",
  "max_tokens": 24576,
  "max_input_tokens": 237568,
  "temperature": 0.6,
  "top_p": 0.95,
  "timeout": 600,
  "extra_body": {
    "chat_template_kwargs": { "enable_thinking": true }
  }
}

{
  "model": "vllm/qwen3.6-27b-fp8",
  "api_base": "http://<VLLM_HOST>:18000/v1",
  "custom_llm_provider": "hosted_vllm",
  "max_tokens": 32768,
  "max_input_tokens": 229376,
  "temperature": 1.0,
  "top_p": 0.95,
  "timeout": 900,
  "stream_timeout": 180,
  "extra_body": {
    "chat_template_kwargs": { "enable_thinking": true }
  }
}

完整 6 個 alias 的 JSON 見 repo 內 configs/litellm/。

Claude Code `.zshrc`

claude-qwen3.6-27b() {
    unset ANTHROPIC_API_KEY OPENAI_BASE_URL OPENAI_API_KEY
    unset AWS_BEARER_TOKEN_BEDROCK CLAUDE_CODE_USE_BEDROCK CLAUDE_CODE_USE_VERTEX

    export ANTHROPIC_BASE_URL="${LITELLM_BASE_URL}"
    export ANTHROPIC_API_KEY="${LITELLM_API_KEY}"
    export ANTHROPIC_AUTH_TOKEN="${LITELLM_API_KEY}"

    export ANTHROPIC_MODEL="qwen3.6-27b-code-act"
    export ANTHROPIC_SMALL_FAST_MODEL="qwen3.6-27b-fast"
    export ANTHROPIC_DEFAULT_OPUS_MODEL="qwen3.6-27b-code-think"
    export ANTHROPIC_DEFAULT_SONNET_MODEL="qwen3.6-27b-code-act"
    export ANTHROPIC_DEFAULT_HAIKU_MODEL="qwen3.6-27b-fast"

    export API_TIMEOUT_MS=900000
    export CLAUDE_CODE_MAX_OUTPUT_TOKENS=16384
    export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1

    claude "$@"
}

alias cct='ANTHROPIC_MODEL=qwen3.6-27b-code-think CLAUDE_CODE_MAX_OUTPUT_TOKENS=24576 claude'
alias ccs='ANTHROPIC_MODEL=qwen3.6-27b-stable     CLAUDE_CODE_MAX_OUTPUT_TOKENS=16384 claude'
alias ccf='ANTHROPIC_MODEL=qwen3.6-27b-fast       CLAUDE_CODE_MAX_OUTPUT_TOKENS=4096  claude'

部署 Checklist

LiteLLM UI 建立 6 個 alias,參數對齊上表
確認 LITELLM_BASE_URL / LITELLM_API_KEY 已設於 shell env
.zshrc 加入 function 與三個 alias
反向代理 proxy_read_timeout ≥ 900s(支援 report)
測試 claude-qwen3.6-27b 能連線、cct / ccs / ccf 切換有效
vLLM 服務確認 --max-model-len 262144 已啟用