deep_load

When DeepSeek R1 landed in January 2025, it shipped with no official ComfyUI support. Ollama and API wrappers existed, but for a full week or two this pack from the Chinese ziwang-com group was about the only thing that ran a real R1 inside the ComfyUI graph. deep_load is the boring half of that bridge: the loader. If you're reading this because you want a reasoner living in your image workflow without standing up a separate service, this is the node that makes that possible - and it has a catch worth knowing about before you go hunting for it.

What it is

deep_load is the loader of a two-node pack (deep_gen does the generating). Think of it as the CLIP loader of the LLM world: it takes a model directory on disk, loads it with HuggingFace transformers, and hands ComfyUI back a DEEP_MODEL object. Nothing generates anything at this stage. Its entire job is getting the weights into memory the way ComfyUI expects.

How it works

The mechanism is small and honest. On startup the node scans ComfyUI/models/deepseek/ for subfolders that contain a config.json, and each match becomes an entry in the model_name dropdown. When you pick one, it loads with AutoModelForCausalLM.from_pretrained(..., torch_dtype="auto") and wraps the result in ComfyUI's ModelPatcher. That last bit matters: it means the same VRAM management that shuffles your SD checkpoints around on smaller cards handles the LLM too, instead of the model eating your GPU and never letting go.

The one input that matters

model_name - the only input. An enum that auto-populates from the models/deepseek/ folder, so there's nothing to type. If a model isn't in the list, it isn't in the folder (or it's not in HF-folder format).

The output is model (DEEP_MODEL), and it wires straight into deep_gen.

Where the model goes

deep_load expects HuggingFace-format model folders, not a single .gguf or a bare safetensors file. Drop the folder in like this:

ComfyUI/models/deepseek/DeepSeek-R1-Distill-Qwen-7B/

The author tested these and calls them fine: DeepSeek-R1-Distill-Qwen-1.5B, -7B, -14B, and DeepSeek-R1-Distill-Llama-8B, on a 24GB 3090. Note the word "Distill" everywhere. The full 671B DeepSeek-R1 is a multi-hundred-GB monster with 163 files on HuggingFace and no realistic chance of fitting on a consumer card - people who tried it on 24GB VRAM / 128GB RAM got told, politely, to go find a distill. The distills are the ones that actually run here.

Installing

Here's the catch the title warned about: the GitHub repo 404s as of mid-2026. github.com/ziwang-com/comfyui-deepseek-r1 is gone (the author's whole ziwang-com account went dark), and the pack was never in the ComfyUI-Manager registry to begin with, so you can't install it by searching Manager. The README's original instruction is still the real one:

cd ComfyUI/custom_nodes
git clone https://github.com/ziwang-com/comfyui-deepseek-r1

…which won't work anymore. Your realistic paths today:

The author's pre-bundled "lazy pack" - a full ComfyUI distribution with a 1.5B model already inside - is still on HuggingFace as zwpython/comfyui-deepseek-r1-portable.
A Wayback Machine copy of azw_nodes.py (the whole node is one file plus a one-line requirements.txt: transformers>=4.37.0).

Troubleshooting

The model_name dropdown is empty. The folder isn't under models/deepseek/, or it lacks config.json. deep_load only lists directories with that file - a lone .safetensors won't show.
It runs but crawls. The v0.2 code pins device_map to CUDA, and the 14B takes ~150 seconds per answer even on a 3090. If your card is smaller, stick with the 7B or 1.5B.
The system prompt field you expected isn't there. v0.2 removed it; that's a deep_gen concern, not this node's.

This pack was a genuine first - a local reasoner inside ComfyUI's own memory manager - and the pattern it set is what a bunch of later LLM nodes copied. It's a period piece now, but if you've got a copy, it still works fully offline.