Please Add Multimodal Image/Video Input Support for Qwen3.8 Models #321

I strongly recommend adding multimodal image and video input support for Qwen3.8 models, especially for models such as:

• Qwen3.8-27B
• Qwen3.8-27B-FP8
• RadixArk/Qwen3.8-27B-NVFP4

At the moment, FreeToken can successfully serve these checkpoints for text inference, but multimodal checkpoints are effectively served as text-only models.

This is currently one of the biggest limitations preventing me from using FreeToken as my primary local inference backend.

Why this matters

Qwen3.8-27B is natively a multimodal model and supports:

• Text
• Images
• Video
• Mixed text-image inputs
• Document / screenshot understanding
• Charts and scientific figures
• Vision-based agent workflows

The multimodal capability is an important part of the model itself, rather than an optional auxiliary feature.

For example, RadixArk/Qwen3.8-27B-NVFP4 still contains the vision-related components and preprocessing configuration, so the checkpoint itself has not been converted into a text-only model.

However, when serving it with FreeToken:

ft serve \
  --model RadixArk/Qwen3.8-27B-NVFP4 \
  --model-source modelscope

text generation works well, but image inputs cannot currently be processed.

Requested feature

It would be extremely useful if FreeToken could add native multimodal support for Qwen3.8, including:

1. Image input support through the OpenAI-compatible API.
2. Support for image_url, local images, and Base64 encoded images.
3. Correct invocation of the Qwen vision encoder / multimodal processor.
4. Mixed text + image conversations.
5. Video input support in the future.
6. Compatibility with quantized checkpoints such as NVFP4 and FP8 whenever the checkpoint preserves the vision module.
7. Multimodal support for /v1/chat/completions and ideally /v1/responses.

For example, I would like to be able to send requests similar to:

{
  "model": "RadixArk/Qwen3.8-27B-NVFP4",
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "Please analyze this image."
        },
        {
          "type": "image_url",
          "image_url": {
            "url": "data:image/jpeg;base64,..."
          }
        }
      ]
    }
  ]
}

Why I think this should be a high-priority feature

FreeToken is already extremely attractive for running recent quantized models locally, especially on consumer NVIDIA GPUs.

However, models such as Qwen3.8 are increasingly designed as unified multimodal models, and serving them as text-only models leaves a significant part of their capabilities unused.

For local research, coding, document analysis, scientific figure understanding, screenshot analysis, and agent workflows, image input is becoming almost as important as text inference.

I am currently running Qwen3.8-27B locally with FreeToken and would very much prefer to continue using FreeToken instead of switching to another inference backend solely because multimodal input is unavailable.

So I strongly hope that Qwen3.8 multimodal support can be prioritized on the roadmap.

Thank you for the excellent project.