GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Rom3rk
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Rom3rk
Model Tree
Qwen/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
text
ComfyUI request-> interpret workflow intent and constraints-> generate, edit, repair, audit, or explain-> return workflow JSON or a conversational answer
Install the required libraries:
bash
pip install --upgrade transformers accelerate safetensors
python
import torchfrom transformers import AutoModelForImageTextToText, AutoProcessorMODEL_ID = "Rom3rk/Qwen3.8-27B-ComfyUI"ENABLE_THINKING = Falseprocessor = AutoProcessor.from_pretrained(MODEL_ID)model = AutoModelForImageTextToText.from_pretrained(MODEL_ID,dtype=torch.bfloat16,device_map="auto",).eval()messages = [{"role": "user","content": [{"type": "text","text": ("Create a ComfyUI UI workflow for an SDXL image-to-image ""pipeline with a ControlNet depth branch and return the ""complete loadable UI workflow JSON."),}],}]inputs = processor.apply_chat_template(messages,tokenize=True,add_generation_prompt=True,enable_thinking=ENABLE_THINKING,return_dict=True,return_tensors="pt",).to(model.device)with torch.inference_mode():generated_ids = model.generate(**inputs,max_new_tokens=4096,do_sample=True,temperature=0.7,top_p=0.8,top_k=20,)new_tokens = generated_ids[:, inputs.input_ids.shape[1]:]response = processor.batch_decode(new_tokens,skip_special_tokens=True,clean_up_tokenization_spaces=False,)[0]print(response)
The model uses Qwen's native chat template. Set ENABLE_THINKING according to the task and inference runtime rather than replacing the template.
text
Create a ComfyUI UI workflow for an SDXL image-to-image pipeline with a ControlNet depth branch, a two-stage sampler, and a final image-save node. Return the complete loadable UI workflow JSON.
text
Modify the following ComfyUI UI workflow to add a second ControlNet branch. Preserve the existing model, sampler settings, node positions, and unrelated links.
text
Repair the invalid links and missing required inputs in this ComfyUI UI workflow. Keep all valid nodes and settings unchanged, and return the corrected workflow JSON.
text
Audit this ComfyUI UI workflow for invalid endpoints, disconnected nodes, incompatible inputs, redundant components, and settings that conflict with the requested generation pipeline.
| Mode | Temperature | Top-p | Top-k | Presence penalty |
|---|---|---|---|---|
| Thinking | 1.0 | 0.95 | 20 | 0.0 |
| Non-thinking | 0.7 | 0.80 | 20 | 1.5 |
These settings follow the recommendations published for Qwen3.8. Availability and parameter names may vary by inference runtime.
For questions, bug reports, workflow compatibility issues, or general feedback, please open a Discussion from this repository's Community tab.
When reporting a workflow issue, please include:
This model is a modified derivative of Qwen/Qwen3.8-27B and is distributed under the Apache License 2.0. See the repository's LICENSE file for the full license text.
Qwen3.8 was developed by the Qwen Team. Model fine-tuning and merging used Unsloth.
bibtex
@misc{qwen38,title = {{Qwen3.8-Max}: A New Bar for Coding and Cowork},author = {{Qwen Team}},month = {August},year = {2026},url = {https://qwen.ai/blog?id=qwen3.8}}