GLM语言丨API

Zhipu's GLM family is one of the strongest Chinese-language LLM lines, and the pack's GLMLanguageAPI is how you get it into a workflow - for prompt rewriting, structured text generation, or any chat-shaped job that a diffusion pipeline needs done in words. Like its DeepSeek sibling in the same pack, it lets you pick the host: 智谱AI (Zhipu's own API) or 硅基流动 (SiliconFlow). Same node, different endpoints, different keys.

It's built on ComfyUI-QING's shared API framework, so it behaves exactly like the other language nodes: multi-turn history in memory, a platform that decides the endpoint and key, friendly model names mapped to each host's real IDs. If you've already configured another QING API node, the setup muscle memory transfers completely.

How it works. The platform dropdown routes between Zhipu's official API and SiliconFlow. Zhipu gets you the full GLM lineup - GLM-4.5, GLM-4.5-Flash, GLM-4, GLM-4-Long, GLM-4-FlashX, and others - while SiliconFlow serves its own set including GLM-4.5-Air, GLM-Z1-32B-0414, and the 9B variants. The tooltip on the model field spells out the reasoning: Flash models for fast daily chat, Long for huge context, the Z-series for reasoning-heavy work. The framework handles the key lookup, applies per-platform token caps, and keeps history rounds of conversation so a multi-step generation can build on itself.

The inputs that matter.

• text_input - the message you're sending (multiline).
• platform - 智谱AI or 硅基流动; match it to the key you hold.
• model - a friendly name from the dropdown (19 options; GLM-4-Plus is the default).
• max_tokens (default 4096) and history (default 10 rounds).
• Optional: temperature, top_p, repetition_penalty, and clear_history.

Outputs: generated_text, conversation_info, total_tokens. That's the standard trio - text out, usage tracking out, and a status/meta string.

How to install. ComfyUI-QING installs as one pack - 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

Set the key in Settings → QING → API Configuration (or the matching env var), restart, and you're live. The README's clone URL is typo'd (GAOSHI-QING, missing the H) - use the URL above or Manager.

Troubleshooting. Key/platform mismatch is the number one error: a Zhipu key with the SiliconFlow platform selected gets you a 401, so keep the dropdown aligned with the key you actually configured. Second, some GLM variants exist on only one host (the tooltip flags GLM-Z1-32B-0414 as a SiliconFlow listing) - if the call fails with a model-not-found, that's usually why. And remember conversation state is per-instance memory: restarts reset it, and clear_history is the manual reset if a long chat starts echoing its own context. As with any API node, your input text leaves the machine to the provider - the standard tradeoff, fine for prompt work.