Qwen3-VL-8B-Thinking

Qwen3-VL-8B-Thinking is an 8-billion-parameter model designed for the in-depth analysis of complex multimodal tasks. Its key feature is the ability for extended, step-by-step reasoning, which is displayed between special tags. This not only leads to a more accurate and well-founded answer but also makes the model's train of thought transparent to the user. Such thorough analysis logically requires more time compared to the faster, but more direct, Instruct version. Architecturally, the model inherits all the innovations of Qwen3-VL: Interleaved-MRoPE for enhanced video understanding, DeepStack for multi-level fusion of visual features, and Text-Timestamp Alignment for precise temporal localization. The context window is 256K tokens with the possibility of expansion up to 1M, and the recommended output sequence length is increased to 40,960 tokens (compared to 16,384 in the Instruct version) to provide sufficient space for extended reasoning chains.

The model achieves strong results on mathematical reasoning benchmarks: 81.4% on MathVista (mini version) and 62.7% on MATH-Vision, outperforming many significantly larger models. At the same time, the model is unmatched on all major 2D/3D Grounding and General VQA benchmarks.

Qwen3-VL-8B-Thinking is particularly effective in scenarios involving the intelligent processing of complex documents, where not only text recognition is required but also an understanding of logical connections, extraction of insights, and drawing conclusions. Advanced OCR capabilities in 32 languages, combined with reasoning mechanisms, make the model an ideal tool for analyzing multilingual documentation, scientific articles, and technical literature. Overall, the model is applicable to any situation that requires not just solving a problem but also a detailed explanation of the solution process.

Announce Date: 15.10.2025
Parameters: 9B
Context: 263K
Layers: 36
Attention Type: Full Attention
Developer: Qwen
Transformers Version: 4.57.0.dev0

Use our pre-built public endpoints for free to test inference and explore Qwen3-VL-8B-Thinking capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at support@immers.cloud or send your request to the sales department at sale@immers.cloud.