Qwen3-VL-8b Image Recognition Issues and Gemma3-27b Mixed Text Image Content

Product Support

AI-generated summary

In a support thread titled Qwen3-VL-8b Image Recognition Issues and Gemma3-27b Mixed Text Image Content, Ivan_Rapekas reports issues with image and PDF handling in LM Studio when using Qwen3-VL-8b and Gemma3-27b models. In LM Studio v3.6.0.beta2, image recognition worked regardless of vision enabled setting, but in v2025.12.0-latest, enabling vision triggers an API error: “Invalid ‘content’: ‘content’ objects must have a ‘type’ field that is either ‘text’ or ‘image_url’.” sam clarifies that LM Studio’s API currently only supports image/text input, not PDFs — a limitation documented in their Python API docs. Ivan_Rapekas acknowledges this, marking the issue as solved for now, and notes that LM Studio’s upcoming v0.4.0 with a new REST API may include PDF support. Separately, Ivan_Rapekas also observes that Gemma3-27b fails to handle mixed text+image content from PDFs, instead generating a fake image URL instead of extracting OCR text. The thread highlights current limitations in multimodal model support within LM Studio, especially around PDF handling, and anticipates future improvements.

1

Hello, I found a topic Managing Images in AI context. I would like to know more about this context.

Could someone clarify current logic of understanding images?

  1. I use Qwen3-VL-8b with LM Studio with OpenAI-compatible API. The hint below says that images are supported by Anthropic, Google and OpenAI models. No chance for Qwen, right?

  2. Qwen3-VL-8b New confusing message when the model cannot recognize a picture/document.

In 3.6.0.beta2:

image
image625×356 32.8 KB

Both in case vision enabled = true and vision enabled = false AI-bot handles the request of image recognition correctly, without any exception.

In v2025.12.0-latest: new option allowed attachments

image
image625×438 38.4 KB

Now with vision enabled = true in returns an error in the dialog:

{“error”:“Invalid ‘content’: ‘content’ objects must have a ‘type’ field that is either ‘text’ or ‘image_url’.”}
  1. Gemma3-27b. Some thoughts about recognizing mixed text+image content. The response currently supports text only. When I ask the model to provide a text from OCR-layer of PDF with separated images, it returns

image
image659×301 34 KB

There is nothing at this URL, the model made a fake link.

Thanks!

2

lmstudio does not have PDF support in completion or responses api.

It only supports image/text from what I can tell.

3

Thank you for the reply! I will mark it as solved and leave a comment here that it was correct for LM Studio 0.3.x. Studio team is currently working on version 0.4.0 with new REST. Hope they add PDF support in their responses.