Qwen3-0.6B-VL

Qwen3-0.6B with eyes. A 0.6B vision-language model built on Qwen3-0.6B and the SigLIP image encoder, trained so that the language model keeps what it could do before: the same chat template, thinking / non-thinking modes, tool calling, and text skills on par with the original Qwen3-0.6B.

• Describe images, answer questions about them, hold multi-turn conversations about a picture.
• Works with or without images, with or without thinking — the text-only behaviour matches Qwen3-0.6B.
• Plain transformers: AutoModelForImageTextToText / pipeline("image-text-to-text"), no trust_remote_code.
• Small enough to run on CPU.
• Try it in your browser, no installation: demo Space.
• Ollama / LM Studio / llama.cpp: GGUF version — ollama run hf.co/oOundefinedOo/Qwen3-0.6B-Visual-GGUF:Q4_K_M.

This is a community model, not an official Qwen release.

Model Overview

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

Quickstart

Tested with transformers==5.18.

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 = 0
thinking_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:

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"])

Tips

• Thinking with images. On open questions ("what is happening here?") the model reasons; on simple visual questions ("is there a dog?", "how many people?") it usually answers directly, which is more accurate for a model this size.
• Tool calling uses the Qwen3 format (tools= in apply_chat_template) and is noticeably more reliable with enable_thinking=True.
• Default sampling follows Qwen3: temperature=0.6, top_p=0.95, top_k=20.

Evaluation

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

Training

1. Alignment: only the projector is trained on ~435k image–caption pairs (vision encoder and LLM frozen).
2. Instruction tuning: projector + LoRA (r=64, all LLM linear layers), merged afterwards. The mix: LLaVA-Instruct and VQA (COCO), conversational yes/no questions built from COCO annotations, image reasoning traces from Qwen3-VL-2B-Thinking, and self-distilled text data — answers of the original Qwen3-0.6B (thinking and non-thinking, multi-turn, tool calling) that keep the language model from drifting.

Limitations

• 224×224 input: small text and fine details are not visible; no OCR, documents, charts, video or grounding.
• Trained on one image per conversation.
• Like any small VLM, it can describe details that are not in the image.

License

Apache 2.0, same as Qwen3-0.6B and SigLIP.