ollama - 💡(How to fix) Fix qwen3.6:35b vision - empty content response on specific pages with long prompt

Background

We are processing a 43-page medical PDF by converting each page to an image (PNG, DPI 300, 3508x2481px) and sending it to a vision model via Ollama for content description and extraction.

We tested two models: qwen3.6:35b and qwen3-vl:30b. Both models process most pages successfully (~30 min per page), but specific pages consistently cause issues depending on the model and prompt used.

Environment

• Ollama version: (ollama --version)
• OS: Ubuntu
• GPU: 2x NVIDIA (~49GB VRAM each)
• Models tested: qwen3.6:35b, qwen3-vl:30b

Problem 1: qwen3.6:35b - Empty content on specific pages

When processing PDF pages converted to images using qwen3.6:35b vision, specific pages consistently return empty content (content='') while eval_count shows ~29,000 tokens generated. The empty response takes ~1 hour compared to ~5 minutes for normal pages.

Key Observations (qwen3.6:35b)

• Same page ALWAYS returns content='' with the same prompt
• Changing the prompt slightly causes DIFFERENT pages to return content=''
• eval_count=29727, done_reason='stop', content=''
• think=False option has no effect
• num_ctx=32768 has no effect
• keep_alive=0 has no effect

Problem 2: qwen3-vl:30b - Infinite hang on specific page

When processing 43 pages of PDF images using qwen3-vl:30b with the same prompt, page 34 causes the model to hang indefinitely (10+ hours with no response). Restarting Ollama server does not resolve the issue — page 34 always hangs.

Key Observations (qwen3-vl:30b)

• All other 42 pages process normally (~30 min each)
• Page 34 never returns a response regardless of wait time
• Restarting Ollama server and retrying produces the same hang
• Same page 34 with a shorter/different prompt may process normally

Reproduction Code

IMAGE_DESCRIPTION_PROMPT = """You are analyzing a page from NCCN clinical guidelines.
Identify and describe ALL content present on this page.

Ignore header (NCCN logo, title, version, copyright) and footer (recommendation note, page number) completely.

Superscripts appear as SMALLER, RAISED characters after a word. Mark them as ※[X].
Example: "Chemotherapy ※[x,y]", "ALL ※[b,c,d,e,f,g]", "blinatumomab ※[aa,bb,cc]"

If flowchart: describe all nodes, arrows, decision branches, and outcomes.
If table: output as markdown table with ※[X] markers.
If footnotes: [footnote X]: <exact text>
If references: [ref X]: <exact text>
If plain text: reproduce exactly with ※[X] markers and section headers.

Preserve all medical terminology exactly. Never omit superscripts."""

response = ollama.chat(
    model="qwen3.6:35b",  # or "qwen3-vl:30b"
    messages=[{
        "role": "user",
        "content": IMAGE_DESCRIPTION_PROMPT,
        "images": [img_b64]  # PNG, 3508x2481px, DPI300
    }],
    options={
        "temperature": 0.0,
        "think": False
    },
    keep_alive=0,
)



## Questions
1. When running the same problematic page as a standalone single image (not as part of a batch),
   the issue still occurs. Is this a known bug in Ollama?

2. Is this a model-specific issue with qwen3.6:35b and qwen3-vl:30b vision,
   or is it related to the Ollama server itself?

3. Is there any workaround or fix available for these issues?