GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
oOundefinedOo
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
oOundefinedOo
Model Tree
Qwen/Qwen3-0.6B
Input Modalities
Output Modalities
Supported Functionality
| Architecture | LLaVA (LlavaForConditionalGeneration): SigLIP-base-patch16-224 → 2-layer MLP projector → Qwen3-0.6B |
| Parameters | ~0.7B (LLM 0.6B + vision 0.09B) |
| Image input | one 224×224 image → 196 image tokens |
| Chat template | Qwen3 (text-only prompts render byte-identically to Qwen3-0.6B), images go first in a message |
Tested with transformers==5.18.
python
from transformers import AutoModelForImageTextToText, AutoProcessor model_name = "oOundefinedOo/Qwen3-0.6B-VL" model = AutoModelForImageTextToText.from_pretrained(model_name, dtype="auto", device_map="auto")processor = AutoProcessor.from_pretrained(model_name) messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"}, {"type": "text", "text": "Describe this image."}, ], }] inputs = processor.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt", enable_thinking=False, # True switches on the thinking mode, as in Qwen3).to(model.device) generated_ids = model.generate(**inputs, max_new_tokens=512)output_ids = generated_ids[0][len(inputs["input_ids"][0]):].tolist() # parse the thinking content (empty when thinking is off)try: index = len(output_ids) - output_ids[::-1].index(151668) # rindex of </think>except ValueError: index = 0thinking_content = processor.decode(output_ids[:index], skip_special_tokens=True).strip("\n")content = processor.decode(output_ids[index:], skip_special_tokens=True).strip("\n")print("thinking content:", thinking_content)print("content:", content)Text-only prompts work exactly as with Qwen3-0.6B — just pass a string as the message content.
The shortest way:
python
from transformers import pipeline pipe = pipeline("image-text-to-text", model="oOundefinedOo/Qwen3-0.6B-VL")messages = [{"role": "user", "content": [ {"type": "image", "url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"}, {"type": "text", "text": "What is in this image?"},]}]print(pipe(text=messages, max_new_tokens=128, return_full_text=False)[0]["generated_text"])tools= in apply_chat_template) and is noticeably more reliable with enable_thinking=True.temperature=0.6, top_p=0.95, top_k=20.All numbers were measured with our own scripts under identical settings; they are not directly comparable to numbers reported by other model cards.
Vision
| Benchmark | Score |
|---|---|
| POPE, F1 (random / popular / adversarial) | 81.6 / 80.7 / 77.5 |
| POPE, F1 with the LLaVA-1.5 short-answer prompt | 80.9 / 84.1 / 76.8 |
| MMBench-EN dev (single pass, not CircularEval) | 60.9 |
POPE images were excluded from training.
Text: does it still behave like Qwen3-0.6B?
| Benchmark | Qwen3-0.6B | Qwen3-0.6B-VL |
|---|---|---|
| GSM8K (300 problems), non-thinking / thinking | 62.0 / 67.3 | 61.3 / 68.7 |
| MMLU (500 questions, 0-shot) | 43.6 | 43.6 |
| Tool calling (60 held-out hermes-function-calling examples), non-thinking / thinking | 61.7 / 58.3 | 63.3 / 61.7 |
Apache 2.0, same as Qwen3-0.6B and SigLIP.
from transformers import AutoModelForImageTextToText, AutoProcessor model_name = "oOundefinedOo/Qwen3-0.6B-VL" model = AutoModelForImageTextToText.from_pretrained(model_name, dtype="auto", device_map="auto")processor = AutoProcessor.from_pretrained(model_name) messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"}, {"type": "text", "text": "Describe this image."}, ], }] inputs = processor.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt", enable_thinking=False, # True switches on the thinking mode, as in Qwen3).to(model.device) generated_ids = model.generate(**inputs, max_new_tokens=512)output_ids = generated_ids[0][len(inputs["input_ids"][0]):].tolist() # parse the thinking content (empty when thinking is off)try: index = len(output_ids) - output_ids[::-1].index(151668) # rindex of </think>except ValueError: index = 0thinking_content = processor.decode(output_ids[:index], skip_special_tokens=True).strip("\n")content = processor.decode(output_ids[index:], skip_special_tokens=True).strip("\n")print("thinking content:", thinking_content)print("content:", content)from transformers import pipeline pipe = pipeline("image-text-to-text", model="oOundefinedOo/Qwen3-0.6B-VL")messages = [{"role": "user", "content": [ {"type": "image", "url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"}, {"type": "text", "text": "What is in this image?"},]}]print(pipe(text=messages, max_new_tokens=128, return_full_text=False)[0]["generated_text"])