Qwen3.5 - A new generation of native multimodal large-scale models open sourced by Alibaba's Tongyi Qianwen.

What is Qwen3.5?

Qwen3.5 is a new generation of native multimodal large-scale model officially released and open-sourced by Alibaba's Tongyi Qianwen team. The first open-source version, Qwen3.5-397B-A17B, adopts an innovative hybrid architecture, combining linear attention (Gated Delta Networks) with sparse hybrid experts (MoE). It boasts a total of 397 billion parameters, but only activates 17 billion parameters per inference, achieving "small but mighty" performance—surpassing the trillion-parameter Qwen3-Max, reducing GPU memory deployment by 60%, and increasing inference throughput by up to 19 times. As a native vision-language model, it unifies text, image, and video tasks through early text-visual fusion and M-RoPE and 3D positional encoding technologies, expanding language support from 119 to 201 languages. In multiple authoritative benchmarks, Qwen3.5 achieved a score of 87.8 in MMLU-Pro, surpassing GPT-5.2; a score of 88.4 in GPQA, exceeding Claude 4.5; and a record-breaking IFBench instruction compliance score of 76.5. The API is priced at only 0.8 yuan per million tokens, which is 1/18 of the price of Gemini 3 Pro. Developers can obtain model weights and interfaces through the Moda Community, HuggingFace, or Alibaba Cloud Bailian.

Main functions of Qwen3.5

• Native multimodal understandingIt unifies the processing of text, images, and videos, and enables cross-modal reasoning through early vision-language fusion.
• Intelligent agent capabilitiesIt supports tool invocation, task planning and execution, and performs excellently in Agent evaluations such as BFCL-V4.
• High-efficiency inference architectureIt adopts a hybrid design of Gated DeltaNet+MoE, and only 17B of the 397B parameters need to be activated to achieve top performance.
• Long context processingIt supports a maximum context length of 1M, and the decoding throughput in 256K scenarios is 19 times that of the previous generation.
• Multilingual supportIt covers 201 languages and dialects, expands the vocabulary to 250,000 words, and improves coding efficiency by 10-60%.
• Code generationIt possesses strong programming capabilities, supporting multi-language code writing, debugging, and complex project development.
• Visual recognitionFine-grained image understanding, object detection, OCR document parsing, and spatial relationship reasoning.
• Search EnhancementIt integrates search agent capabilities such as BrowseComp to enable online information retrieval and integration.

Technical Principles of Qwen3.5

• Hybrid attention architectureIt integrates Gated DeltaNet linear attention and Gated Attention sparse attention to achieve adaptive allocation of computational resources.
• MoE Sparse ActivationOf the total 397B parameters, only 17B (approximately 4.3%) are activated. Inference costs are reduced by using one shared expert and 512 routing experts (10 are activated each time).
• Gating mechanism optimizationThe paper, which uses head-specific sigmoid gating instead of traditional softmax attention, won the 2025 NeurIPS Best Paper Award.
• Native multimodal fusionBy using M-RoPE and 3D position coding to natively embed spatiotemporal information into LLM, early vision-language fusion can be achieved.
• Heterogeneous training frameworkThe strategy of decoupling and parallelizing the vision and language components, and the sparse activation to achieve cross-module computation overlap, result in training throughput close to the baseline of pure text.
• FP8 Mixed PrecisionThe native FP8 pipeline uses low precision for activation, MoE routing, and GEMM, while maintaining BF16 for the sensitive layer and reducing video memory usage by more than 50%.
• Asynchronous reinforcement learningBuild a scalable asynchronous RL framework to support full-size models, achieve training-push separation and dynamic load balancing, and accelerate end-to-end by 3-5 times.

Qwen3.5 project address

• Hugging Face Model Libraryhttps://huggingface.co/collections/Qwen/qwen35
• GitHub repository: https://github.com/QwenLM/Qwen3.5

How to use Qwen 3.5

• Web-based experienceVisit Qwen Chat or download the Qwen APP/PC client to directly enter text or upload pictures/videos for multimodal interaction.
• API calls: Obtain API keys through Alibaba Cloud's Bailian platform and access model services at a low price of 0.8 yuan per million tokens.
• Local deploymentDownload the open-source weight Qwen3.5-397B-A17B from HuggingFace or the Moda community and deploy and run it in a local or private cloud environment.
• Intelligent agent development: Utilize the model's native agent capabilities and combine them with tool API calls to develop applications for automated task execution, data analysis, and other applications.
• Multilingual applicationsYou can directly input any of the 201 supported languages to start a conversation without needing to configure an additional translation module.
• Long document processingUpload extremely long texts, PDFs, or videos, and the model will automatically handle complex inference tasks with contexts up to 1MB.

Qwen3.5 model pricing

• Enter priceEach million tokens costs only 0.8 yuan, making it one of the lowest prices among the world's top models.
• Price comparisonIt is only 1/18 the size of the Gemini 3 Pro and 1/15 the size of the GPT-5.2, offering a significant price-performance advantage.
• Open source and freeThe weights of the Qwen3.5-397B-A17B model are completely open source; developers can download and deploy them locally for free.

Application Scenarios of Qwen3.5

• Intelligent Customer ServiceSupports multi-turn dialogues and complex question answering in 201 languages, suitable for global enterprise customer service systems.
• Content creationIt assists in copywriting, video script generation, multilingual translation, and localization of content.
• Code developmentIt provides code generation, debugging, and refactoring services to support the development and maintenance of complex software projects.
• Visual analysisIt enables image recognition, document OCR, video content understanding, and fine-grained visual question answering.
• Intelligent agent automationBuild automated workflows to perform agent tasks such as data analysis, information retrieval, and task scheduling.
• Educational SupportIt supports multilingual teaching, homework correction, personalized learning path planning, and knowledge Q&A.