DeepSeek语言丨API

DeepSeek has a reputation in this ecosystem for exactly what the tooltip says: code generation, math, and logical reasoning. DeepSeekLanguageAPI puts that inside your ComfyUI graph as a chat node - you type a question, it returns a generated answer, and because the pack keeps a multi-turn history you can actually hold a conversation with it across runs. Where it gets interesting is the platform dropdown: DeepSeek isn't one endpoint, it's a model that multiple Chinese providers host, and this node lets you pick who serves it - Volcano Engine (火山引擎), Alibaba Cloud Bailian (阿里云百炼), SiliconFlow (硅基流动), or Tencent Cloud (腾讯云).

It's part of ComfyUI-QING's API category, built on the pack's shared API framework. The whole "one key per provider" setup lives in one place, which is the pattern the KB's external-API-nodes essay praises: keys are stored in settings, not pasted into every node.

How it works. Behind the scenes, each platform option maps to a base URL, an environment-variable key name, and a translation table that converts the friendly model names you see in the dropdown into each provider's real API model IDs. So DeepSeek-R1 becomes deepseek-r1-250528 on Volcano, deepseek-r1-0528 on Bailian, and deepseek-ai/DeepSeek-R1 on SiliconFlow. The framework grabs your key (env var like VOLCENGINE_API_KEY, DASHSCOPE_API_KEY, SILICONFLOW_API_KEY, or TENCENT_LKEAP_API_KEY, or the QING settings panel), applies token limits per platform, and manages conversation history in memory.

The inputs that matter.

• text_input - your prompt (multiline).
• platform - which provider hosts the model. This decides the key and the endpoint, so pick the one you actually have a key for.
• model - DeepSeek-R1, V3.1, V3.1-Terminus, or V3 (Terminus is SiliconFlow-only, per the tooltip).
• max_tokens (default 4096) and history (default 20 rounds of retained conversation).
• Optional: temperature, top_p, repetition_penalty (defaults 0.7 / 0.9 / 1.1), and clear_history to reset the conversation.

Outputs: generated_text, conversation_info, and total_tokens. Wire generated_text into a prompt template or a text preview.

How to install. ComfyUI-QING is a one-pack install - search "ComfyUI-QING" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/GAO-SHIQING/ComfyUI-QING
cd ComfyUI-QING
python install_dependencies.py

Then set your key: Settings → QING → API Configuration, or an environment variable. Restart after installing. The README's clone URL is typo'd (GAOSHI-QING, missing the H) - use the URL above or Manager.

Troubleshooting. The most common failure is a key/endpoint mismatch - you picked 硅基流动 in the dropdown but only have a Bailian key, so the call 401s. Match the platform to the key you actually hold. Second: the model names are friendly labels, not raw IDs - don't try to type a provider's internal model string into the field. Third: conversation state lives in memory per node instance, so a ComfyUI restart clears it - use clear_history if the conversation is drifting but you want to keep the graph. And this being an API node, your text leaves the machine to the provider - fine for prompt work, worth remembering if you're feeding it workflow internals you'd rather not share.