Alibaba’s Qwen AI Releases Compact Dense Qwen3-VL 4B/8B (Instruct & Thinking) With FP8 Checkpoints

The recent release of Alibaba’s Qwen3-VL models marks a significant advancement in the realm of AI and business management. This update introduces compact and dense models at 4B and 8B scales, designed to operate efficiently with low VRAM while maintaining a robust capability surface.

Target Audience Analysis

The primary audience for these models includes AI developers, data scientists, and business managers looking to integrate advanced AI solutions into their operations. Their pain points often revolve around the high resource demands of existing models, the complexity of deployment, and the need for scalable solutions that can handle multimodal tasks effectively.

Goals for this audience typically include:

• Reducing operational costs associated with AI deployment
• Enhancing the efficiency of AI applications in business processes
• Leveraging advanced AI capabilities without extensive infrastructure

Interests include the latest advancements in AI technology, practical applications of multimodal models, and tools that facilitate easier integration into existing systems. Communication preferences lean towards technical documentation, detailed specifications, and case studies demonstrating real-world applications.

Release Overview

Alibaba’s Qwen team has expanded its multimodal lineup with the introduction of the Qwen3-VL models, which include:

• Qwen3-VL-4B
• Qwen3-VL-8B
• Instruct and Thinking editions for both models
• FP8-quantized checkpoints for low-VRAM deployment

These models are designed to complement the previously released 30B and 235B tiers while retaining a comprehensive capability surface that includes:

• Image/video understanding
• OCR in 32 languages
• Spatial grounding
• GUI/agent control

Technical Specifications

The Qwen3-VL models feature a native context length of 256K, expandable to 1M, and maintain the full feature set of their larger counterparts. Key architectural updates include:

• Interleaved-MRoPE for improved positional encoding
• DeepStack for enhanced image-text alignment
• Text–Timestamp Alignment for event localization in video

These enhancements ensure that the models can handle long-document and video comprehension tasks effectively.

FP8 Checkpoints

The FP8 checkpoints utilize fine-grained FP8 quantization with a block size of 128, achieving performance metrics nearly identical to the original BF16 checkpoints. This allows teams to evaluate precision trade-offs on multimodal stacks without the burden of re-quantization and re-validation. However, it is important to note that the current tooling status indicates that Transformers do not yet support loading these FP8 weights directly; users are recommended to utilize vLLM or SGLang for serving.

Key Takeaways

In summary, Alibaba’s Qwen AI has released:

• Dense Qwen3-VL 4B and 8B models in both Instruct and Thinking variants
• FP8 checkpoints that support low-VRAM deployment
• A preserved capability surface that includes 256K to 1M context, OCR, spatial grounding, video reasoning, and GUI/agent control
• Model sizes: Qwen3-VL-4B approximately 4.83B parameters; Qwen3-VL-8B-Instruct approximately 8.77B parameters

These developments position Qwen3-VL models as practical solutions for teams targeting deployment on single-GPU or edge budgets.

Further Resources

For additional information, check out the model on Hugging Face. You can also explore our GitHub Page for tutorials, codes, and notebooks. Stay updated by following us on Twitter and join our 100k+ ML community on SubReddit. Additionally, connect with us on Telegram.