Skip to main content

Qwen3-VL 32B Instruct (FP8)

Vision-language model with a 256K native context, FP8-quantized.

Overview​

  • model_name: hf-Qwen-Qwen3-VL-32B-Instruct-FP8 (HuggingFace import, currently free)
  • Source: Qwen/Qwen3-VL-32B-Instruct-FP8
  • Architecture: Qwen3VLForConditionalGeneration — 33B parameters, fine-grained FP8 (block size 128), 64 attention heads
  • Served on: GPU via vLLM. ~33 GB of weights; runs on one large card or tensor-parallel across two.
  • Status: candidate — imported and wired, pending its inference test

When to use​

✅ Reading documents, screenshots, and charts; OCR across 32 languages; visual grounding and spatial questions. It is the largest vision-language model in the catalog.

❌ Text-only work — a text model of the same size is cheaper and faster. For plain chat see Qwen 2.5 7B; for a smaller multimodal option the catalog also serves qwen3-vl-8b.

Live specs​

curl "https://api.colabhive.com/api/builder/v1/endpoints?visibility=public&search=Qwen3-VL-32B"

Quick start​

curl -X POST "https://api.colabhive.com/api/builder/v1/endpoints/<endpoint_id>/infer" \
  -H "X-API-Key: $COLABHIVE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input": {"messages": [{"role": "user", "content": [{"type": "image_url", "image_url": {"url": "https://example.com/receipt.png"}}, {"type": "text", "text": "Read every line item and total."}]}], "max_tokens": 800}}'
First call after idle

Cold-starting 33 GB of weights takes several minutes. The request is queued rather than rejected, and completes once the replica is warm.

Authors: José Luis Minich, Maximiliano Lucius.