Load Qwen Model

Load Qwen Model is the unglamorous foundation node of the whole QwenScope pack. Nothing happens until it's on the canvas, because every capture, inspect and steer node needs a QSCOPE_MODEL wire from it. It loads a Qwen base LM plus its tokenizer into memory so the rest of the pack can poke at its residual stream.

The name is not a lie but it is a clue: this loads base models, not chat models. You're not here for chat, you're here for interpretability - you want the raw autoregressive backbone whose internal activations you can read and nudge. The drop-in default is Qwen/Qwen3.5-2B, which is small enough to fit alongside an image model and fast enough that analyzing a prompt takes seconds.

How it works

Under the hood it's a straightforward HuggingFace load: AutoModelForCausalLM.from_pretrained plus AutoTokenizer, wrapped so ComfyUI sees it as a node. Two things are worth knowing before you wire it up.

First, there's a model cache of exactly one. QwenScope only keeps one loaded LLM in VRAM at a time, so if you have two Load Qwen Model nodes in a graph (say, a 2B and a 9B), loading the second evicts the first. In practice that's fine - you run one analysis model at a time - but it's why swapping a preset can feel like a cold reload.

Second, dtype and device_map both default to auto, which is what you want on a typical GPU box. If you're short on VRAM, bfloat16 and float32 are the honest alternatives; device_map=auto will shard across GPU/CPU if it must, so a 9B can run on a modest card, just slowly.

The inputs that matter

• preset - the dropdown of registered base models (Qwen3 and Qwen3.5 families, MoE variants included). This drives everything else, so pick the one whose SAE you'll pair it with.
• custom_model_id - only used when preset is custom. This is your escape hatch: type any HF repo id and it loads whatever causal LM you point at. Useful when Qwen ships a new family before the registry catches up.
• dtype / device_map - leave at auto unless you know better.

Outputs: model (QSCOPE_MODEL) feeds the Analyze nodes and Generate; info is a plain STRING telling you what loaded, the hidden size, layer count and device - read it once to confirm you got the model you thought you did.

Installing it

ComfyUI Manager is the easy path: search "ComfyUI QwenScope" and install. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Nynxz/ComfyUI-QwenScope

then restart ComfyUI. Dependencies are transformers>=4.45, huggingface_hub, safetensors, Pillow and numpy - ComfyUI already ships torch, and most installs already have transformers and huggingface_hub for other nodes. If pip complains, pip install -e . inside the pack folder pins everything. Model weights download from HuggingFace on first load; if a repo is gated, set HF_TOKEN before launching ComfyUI.

Common issues

The classic stumble is a d_model mismatch further down the graph: if your SAE layer's dictionary expects a 4096-d residual but you loaded a 2B model (2048-d), the Analyze node will throw. Load the base model that matches the SAE pack you picked - the registry is keyed to specific base models for exactly this reason. And remember the first load of any model is slow: it's downloading gigabytes, not hanging.