ComfyUI-Qwen3.5

Custom ComfyUI nodes for the Qwen3.5 family — unified natively multimodal models with image, video, and text understanding.

Four nodes included:

• Qwen 3.5 — transformers-based, supports image + video + text, FP16/8-bit/4-bit quantization
• Qwen 3.5 (GGUF) — llama.cpp-based, 9x faster (152 tok/s vs 17 tok/s), uses GGUF quantized models
• Qwen 3.5 (WaveSpeed API) — cloud API, no local GPU needed, access up to 397B parameter models
• Load Image from URL — utility node to load images from any URL

Features

• Image understanding — describe, analyze, or answer questions about images
• Video understanding — summarize or analyze video content (transformers node)
• Text generation — pure text tasks (reasoning, writing, coding)
• Thinking mode — optional chain-of-thought reasoning before response
• GGUF inference — 152 tokens/second via llama.cpp (Q4_K_XL on RTX PRO 6000)
• Quantization — FP16, 8-bit, 4-bit (transformers) or GGUF quantizations (Q4-Q8, BF16)
• CPU support — both nodes work on CPU (transformers uses FP32, GGUF needs llama.cpp built without CUDA)
• ComfyUI compatible — automatically handles cudaMallocAsync compatibility

Installation

Clone into your ComfyUI custom nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/DanielBartolic/ComfyUI-Qwen3.5.git

Then install dependencies for the node(s) you want to use:

Transformers node only

pip install -r ComfyUI-Qwen3.5/requirements.txt -r ComfyUI-Qwen3.5/requirements-transformers.txt

GGUF node only

pip install -r ComfyUI-Qwen3.5/requirements.txt -r ComfyUI-Qwen3.5/requirements-gguf.txt

The GGUF node requires llama.cpp — see Building llama.cpp below.

Both nodes

pip install -r ComfyUI-Qwen3.5/requirements.txt \
            -r ComfyUI-Qwen3.5/requirements-transformers.txt \
            -r ComfyUI-Qwen3.5/requirements-gguf.txt

Note: The transformers node automatically sets HF_DEACTIVATE_ASYNC_LOAD=1 to prevent OOM errors caused by transformers >= 5.2.0's parallel weight loading conflicting with ComfyUI's cudaMallocAsync allocator. No special flags needed.

---

Model Paths (Manual Placement)

Both nodes look for models inside ComfyUI/models/LLM/. Models are auto-downloaded on first use, but you can place them manually to avoid downloads.

Transformers node

Place the full model folder at:

ComfyUI/models/LLM/<model-name>/

For example, for Qwen3.5-9B:

ComfyUI/models/LLM/Qwen3.5-9B/
├── config.json
├── model-00001-of-00005.safetensors
├── model-00002-of-00005.safetensors
├── ...
├── tokenizer.json
├── tokenizer_config.json
└── ...

The node checks for config.json in that directory — if it exists, no download happens.

You can download the model files manually from HuggingFace:

• Qwen3.5-0.8B → ComfyUI/models/LLM/Qwen3.5-0.8B/
• Qwen3.5-2B → ComfyUI/models/LLM/Qwen3.5-2B/
• Qwen3.5-4B → ComfyUI/models/LLM/Qwen3.5-4B/
• Qwen3.5-9B → ComfyUI/models/LLM/Qwen3.5-9B/
• Qwen3.5-27B → ComfyUI/models/LLM/Qwen3.5-27B/

GGUF node

Place GGUF files at:

ComfyUI/models/LLM/<model-name>-GGUF/

For example, for Qwen3.5-9B Q4_K_XL:

ComfyUI/models/LLM/Qwen3.5-9B-GGUF/
├── Qwen3.5-9B-UD-Q4_K_XL.gguf      ← the model
└── mmproj-BF16.gguf                  ← vision projector (required for image input)

The GGUF filename format is <model>-<quantization>.gguf, with UD- prefix for Unsloth Dynamic quantizations (XL variants).

Download manually from HuggingFace:

• unsloth/Qwen3.5-9B-GGUF — pick your quantization + mmproj-BF16.gguf
• unsloth/Qwen3.5-4B-GGUF
• unsloth/Qwen3.5-2B-GGUF
• unsloth/Qwen3.5-0.8B-GGUF

---

Building llama.cpp

The GGUF node calls llama-mtmd-cli (the multimodal CLI from llama.cpp). You must build it from source.

With CUDA (GPU — recommended)

git clone https://github.com/ggml-org/llama.cpp
cmake llama.cpp -B llama.cpp/build -DGGML_CUDA=ON
cmake --build llama.cpp/build --config Release -j$(nproc)
cp llama.cpp/build/bin/llama-mtmd-cli /usr/local/bin/

CPU only (no CUDA required)

git clone https://github.com/ggml-org/llama.cpp
cmake llama.cpp -B llama.cpp/build
cmake --build llama.cpp/build --config Release -j$(nproc)
cp llama.cpp/build/bin/llama-mtmd-cli /usr/local/bin/

For CPU-only, set n_gpu_layers to 0 in the node settings.

Note: Installing the llama-cpp-python pip package does NOT provide llama-mtmd-cli. You must build llama.cpp from source as shown above.

Verifying the build

After building, confirm the binary works:

llama-mtmd-cli --version

If you get llama-mtmd-cli not found, either:

1. Copy the binary to a directory in your PATH: cp llama.cpp/build/bin/llama-mtmd-cli /usr/local/bin/
2. Or set the cli_path input in the node to the full path: e.g. /home/user/llama.cpp/build/bin/llama-mtmd-cli

The node searches for llama-mtmd-cli in this order:

1. The cli_path input (if set)
2. System PATH
3. /usr/local/bin/llama-mtmd-cli
4. /opt/llama.cpp/build/bin/llama-mtmd-cli
5. /workspace/llama.cpp/build/bin/llama-mtmd-cli

---

Node: Qwen 3.5

Transformers-based node. Found under Qwen3.5 in the node menu. Supports image, video, and text.

Works on both GPU (CUDA) and CPU. On CPU, models load in FP32 (slower but functional).

Inputs

Input	Type	Default	Description
model	dropdown	Qwen3.5-9B	Model size (0.8B / 2B / 4B / 9B / 27B)
prompt	STRING	required	Text prompt for the model
system_prompt	STRING	""	Optional system prompt
max_tokens	INT	4096	Maximum tokens to generate
temperature	FLOAT	1.0	Sampling temperature
top_p	FLOAT	0.95	Nucleus sampling
top_k	INT	20	Top-K sampling
repetition_penalty	FLOAT	1.0	Repeated token penalty
enable_thinking	BOOLEAN	True	Enable chain-of-thought reasoning
quantization	dropdown	FP16	FP16 / 8-bit / 4-bit
keep_model_loaded	BOOLEAN	True	Keep model in VRAM between runs
seed	INT	1	Random seed
image	IMAGE	optional	Single image input
video	IMAGE	optional	Video frames (batch of images)
frame_count	INT	16	Max frames to sample from video

Supported Models

Model	Parameters	VRAM (FP16)	VRAM (8-bit)	VRAM (4-bit)
Qwen3.5-0.8B	0.8B	~2 GB	~1 GB	~1 GB
Qwen3.5-2B	2B	~5 GB	~3 GB	~2 GB
Qwen3.5-4B	4B	~9 GB	~6 GB	~4 GB
Qwen3.5-9B	9.65B	~20 GB	~12 GB	~7 GB
Qwen3.5-27B	27B	~56 GB	~30 GB	~17 GB

---

Node: Qwen 3.5 (GGUF)

llama.cpp-based node. 9x faster than transformers on GPU. Found under Qwen3.5 in the node menu. Supports image and text (no video).

Works on both GPU and CPU. For CPU-only, set n_gpu_layers to 0 and build llama.cpp without -DGGML_CUDA=ON.

Inputs

Input	Type	Default	Description
model	dropdown	Qwen3.5-9B	Model size (0.8B / 2B / 4B / 9B / 27B)
quantization	dropdown	Q4_K_XL	GGUF quantization level
prompt	STRING	required	Text prompt for the model
system_prompt	STRING	""	Optional system prompt
max_tokens	INT	4096	Maximum tokens to generate
temperature	FLOAT	0.7	Sampling temperature
top_p	FLOAT	0.8	Nucleus sampling
top_k	INT	20	Top-K sampling
repeat_penalty	FLOAT	1.0	Repeated token penalty
n_gpu_layers	INT	99	GPU layers to offload (99 = all, 0 = CPU only)
ctx_size	INT	8192	Context window size
enable_thinking	BOOLEAN	False	Enable chain-of-thought reasoning
seed	INT	1	Random seed
image	IMAGE	optional	Image for vision tasks
cli_path	STRING	""	Path to llama-mtmd-cli (auto-detected if empty)

GGUF Quantizations

All from unsloth/Qwen3.5-9B-GGUF:

Quantization	Size	Speed (RTX PRO 6000)
Q4_K_XL (Unsloth Dynamic)	6.0 GB	~152 tok/s
Q4_K_M	5.7 GB	~150 tok/s
Q5_K_XL (Unsloth Dynamic)	6.7 GB	~130 tok/s
Q6_K_XL (Unsloth Dynamic)	8.8 GB	~110 tok/s
Q8_0	9.5 GB	~90 tok/s
BF16 (full precision)	17.9 GB	~60 tok/s

---

Node: Qwen 3.5 (WaveSpeed API)

Cloud-based node using WaveSpeed's OpenAI-compatible API. No local GPU needed — runs on WaveSpeed's infrastructure. Found under Qwen3.5 in the node menu.

Requires a WaveSpeed API key. Set WAVESPEED_API_KEY environment variable or pass it directly in the node.

pip install openai

Inputs

Input	Type	Default	Description
model	dropdown	Qwen3.5-27B	Model size (35B-A3B to 397B-A17B)
prompt	STRING	required	Text prompt
system_prompt	STRING	""	Optional system prompt
max_tokens	INT	4096	Maximum tokens to generate
temperature	FLOAT	1.0	Sampling temperature
top_p	FLOAT	0.95	Nucleus sampling
top_k	INT	20	Top-K sampling
api_key	STRING	""	WaveSpeed API key (or use env var)
image	IMAGE	optional	Image input (resized + base64 encoded)
image_url	STRING	""	Image URL (sent directly, preferred over image input)

Available Models

Model	Pricing (input/output per M tokens)
Qwen3.5-9B	—
Qwen3.5-35B-A3B	$0.16 / $1.30 (cheapest)
Qwen3.5-Flash	—
Qwen3.5-27B	$0.20 / $1.60
Qwen3.5-Plus	—
Qwen3.5-122B-A10B	$0.26 / $2.10
Qwen3.5-397B-A17B	$0.39 / $2.30 (best quality)

---

Node: Load Image from URL

Utility node to download an image from any public URL and output it as a ComfyUI IMAGE tensor. Shows preview after execution. Found under Qwen3.5 in the node menu.

Inputs

Input	Type	Description
url	STRING	Public image URL

Outputs

Output	Type	Description
IMAGE	IMAGE	Downloaded image as tensor
URL	STRING	Pass-through of the input URL

---

Output (Qwen 3.5 nodes)

Output	Type	Description
RESPONSE	STRING	Model's text response (thinking stripped)
THINKING	STRING	Extracted reasoning content (empty if thinking disabled)

Recommended Sampling Parameters

From the Qwen3.5 README:

Mode	Temperature	Top-p	Top-k	Repetition Penalty
Thinking	1.0	0.95	20	1.0
Instruct (default)	0.7	0.8	20	1.0

Troubleshooting

llama-mtmd-cli not found

You need to build llama.cpp from source — see Building llama.cpp. The llama-cpp-python pip package does not include this binary.

ggml_cuda_init: failed to initialize CUDA (Blackwell / newer GPUs)

If cmake reports CMAKE_CUDA_ARCHITECTURES_NATIVE=No CUDA devices found, specify the architecture manually:

# Blackwell (RTX PRO 6000, B200, etc.) — SM_120:
cmake llama.cpp -B llama.cpp/build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=120

# Ada Lovelace (RTX 4090, L40, etc.) — SM_89:
cmake llama.cpp -B llama.cpp/build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=89

# Hopper (H100, H200) — SM_90:
cmake llama.cpp -B llama.cpp/build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=90

This happens in containers where the GPU is available via nvidia-smi but not visible to the CUDA runtime during build.

CPU-only usage

Both nodes work without a GPU:

• Transformers node: Just select a model — it auto-detects CPU and uses FP32. 8-bit/4-bit quantization requires a CUDA GPU.
• GGUF node: Build llama.cpp without -DGGML_CUDA=ON and set n_gpu_layers to 0.

Models not downloading

If auto-download fails, place models manually — see Model Paths above.

Requirements

Transformers node:

• transformers >= 5.2.0
• torch
• bitsandbytes (for quantization, GPU only)
• accelerate

GGUF node:

• llama-mtmd-cli binary (built from llama.cpp source)
• huggingface-hub (for auto model downloads)
• numpy, Pillow, torch

License

Apache-2.0