Lmcq Deep Loader

Most "LLM in my ComfyUI workflow" setups mean calling out to OpenAI or Claude and paying per token. LmcqDeepLoader is the alternative: it loads a locally-downloaded DeepSeek-R1-Distill model straight into the graph so a downstream generation node (LmcqDeepGen, from the same pack) can run entirely offline. If you want prompt expansion, captioning, or any text-generation step inside a workflow without an API key or a network dependency, this is the loader that makes it possible.

What it actually loads. DeepSeek-R1-Distill is a family of open reasoning models distilled down from the full DeepSeek-R1 into smaller, runnable sizes - Qwen-based at 1B/7B/14B/32B, and a Llama-based 70B. This node doesn't download anything for you; it just points at whatever you've already put in models/deepseek/ and hands it off in a format LmcqDeepGen understands.

The one input. model_name - a dropdown reading whatever folders you've put under models/deepseek/. There's nothing else to set here; all the generation parameters (temperature, sampling, prompts) live on the downstream node, not this one.

The output. model, typed DEEP_MODEL - a custom type specific to this pack, so the only thing you can plug it into is LmcqDeepGen. It won't connect to anything expecting a standard MODEL type (a checkpoint output, for instance) - this is a completely separate model family from the image-generation side of ComfyUI.

Getting the model files - this is the part people skip. Unlike most ComfyUI loaders, this one won't auto-fetch anything, and the pack's README is explicit about the manual step: download every file for your chosen size from Hugging Face (deepseek-ai/DeepSeek-R1-Distill-Qwen-1B through -32B, or -Llama-70B for the largest), keep them together, and place the whole folder under models/deepseek/ using the model's name as the folder name - e.g. models/deepseek/DeepSeek-R1-Distill-Qwen-7B/. Grab the whole repo, not just the .safetensors - tokenizer and config files live alongside the weights and the loader needs all of them.

Picking a size. This isn't documented in the pack itself, but it's the obvious practical question: bigger Distill sizes need meaningfully more VRAM/RAM, same as any local LLM. The 1B and 7B variants are the realistic starting point on a normal gaming GPU; 32B and 70B are workstation-class hardware. If you just want a prompt-rewriting step in an image workflow, you don't need the biggest model - start small and go up only if quality is the bottleneck.

Installing the node pack itself. Through ComfyUI Manager (search "ComfyUI-LMCQ") or:

cd ComfyUI/custom_nodes
git clone https://github.com/sebord/ComfyUI-LMCQ

Restart ComfyUI, then handle the model download separately as above - installing the node and having a usable model are two different steps here.

Troubleshooting. If model_name shows an empty dropdown, the loader isn't finding your models/deepseek/ folder or it's empty - double check the folder name matches the model exactly and that all files (not just the weights) made it into it. And the pack-wide caveat applies here too: if none of the Lmcq nodes are showing up at all (not just this one), check your ComfyUI console for a ModuleNotFoundError referencing the pack's compiled runtime-protection module - a known, reported install issue unrelated to DeepSeek specifically, with the fix documented in the README's January 2026 update log.