furiosa-ai/Qwen3-VL-32B-Thinking

VISIONPricing:Input $0.416 / Output $1.664Concurrent Unit Cost:2Model Size:33.4BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 24, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Qwen3-VL-32B-Thinking is a 33.4 billion parameter dense vision-language model from the Qwen3-VL series, developed by Qwen. It integrates a vision encoder with a dense transformer decoder, utilizing Interleaved-MRoPE and DeepStack for multimodal processing. This model excels at visual understanding tasks including OCR, document analysis, spatial reasoning, and video comprehension, and natively supports tool calling. As a "Thinking" edition, it explicitly generates a chain-of-thought before providing its final answer, making it suitable for applications requiring transparent reasoning.

Loading preview...

Overview

Qwen3-VL-32B-Thinking is a 32-billion-parameter dense vision-language model developed by Qwen, part of the Qwen3-VL series. It combines a vision encoder with a dense transformer decoder, leveraging Interleaved-MRoPE positional embeddings and DeepStack multi-level feature fusion to process images, videos, and text. This model is specifically the Thinking edition, designed to emit an explicit chain-of-thought before delivering its final answer, which is parsed separately by Furiosa-LLM.

Key Capabilities

  • Vision-Language Understanding: Handles diverse visual tasks such as OCR, document and chart analysis, spatial reasoning, and video comprehension by accepting OpenAI-style multimodal chat messages with image_url content parts.
  • Reasoning: Generates a transparent chain-of-thought, which Furiosa-LLM can parse using the qwen3 reasoning parser, providing insight into the model's decision-making process.
  • Tool Calling: Supports function calling via the hermes tool-call parser, enabling integration with external tools and APIs.
  • Hardware Optimization: Optimized for running on FuriosaAI RNGD hardware with Furiosa-LLM, utilizing a tensor-parallel strategy of 32 PEs across four RNGD cards.

Usage Considerations

  • The model runs in its original precision without quantization.
  • When serving with Furiosa-LLM, specific flags can control multimodal behavior, such as image limits per prompt, allowed local media paths, and whitelisted media domains.
  • The reasoning field is returned separately from the final answer, allowing developers to access the model's thought process.

For more details, refer to the upstream Qwen/Qwen3-VL-32B-Thinking model card.

Finetunes
1 models