chat-deepseek
GitHub调用 deepseek-v4-flash 模型处理代码读写、重构及多步自动化规划。适用于编码和工具使用场景，性价比高。直接返回纯文本答案，不修改内容，仅在输出包含特定节省提示时附带该信息。
触发场景
安装
npx skills add zerogpu/zerogpu-router --skill chat-deepseek -g -y
SKILL.md
Frontmatter
{
"name": "chat-deepseek",
"description": "Chat with deepseek-v4-flash, a 284B MoE model (13B active per token) with a 1M-token context window, tuned for coding and agentic workflows. Use for reading or writing code across a large codebase, porting and refactoring, or planning multi-step automation. Cheaper than chat-glm at the same context size.",
"allowed-tools": "Bash(zerogpu chat *)",
"argument-hint": "<text> [-i <instructions>]"
}
Call deepseek-v4-flash. $ARGUMENTS is the raw prompt — pass it verbatim, no escaping or quoting required (the heredoc below handles every shell metacharacter, newline, quote, and paren safely):
ZGPU_TEXT=$(cat <<'ZGPU_END_OF_INPUT'
$ARGUMENTS
ZGPU_END_OF_INPUT
)
zerogpu chat "$ZGPU_TEXT" -m deepseek-v4-flash
If the user supplied system instructions, append -i "<instructions>" after the model flag.
At $0.07 / $0.14 per 1M input/output tokens this is the cheaper of the two 1M-context models — roughly a sixteenth of /zerogpu-router:chat-glm. Prefer it whenever the task is code or tool-use rather than sheer input size. For a prompt that fits in 131K tokens, /zerogpu-router:chat is cheaper still.
This model is served by the Chat Completions API rather than the Responses API; the CLI routes it automatically. Output is the assistant's answer as plain text — the model's reasoning trace is omitted, since this skill does not pass the CLI's -r flag. Relay the answer as-is — do not rewrite or expand it.
Savings note: only if the command output literally contains a line starting with 💰 ZeroGPU savings, append that exact line, unchanged, as the last line of your reply. If no such line is present, say nothing about savings and do not mention or suggest /zerogpu-router:cost-savings — this note is intentionally occasional, not shown every time.
版本历史
- cee9321 当前 2026-08-02 21:14
