News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgfnQwen3.5-397B-A17B

Qwen3.5-397B-A17B

A large multimodal language model with sparse activation.

Alibaba Cloud released Qwen3.5-397B-A17B as part of its Qwen3.5 series. This causal language model combines vision and text processing using a hybrid architecture. It features 397 billion total parameters but activates only 17 billion per token, aiming for high efficiency and broad linguistic support.

Share on WhatsApp

  • 397 billion total parameters with 17 billion active per token
  • Supports 201 languages and dialects
  • Native context length of 262,144 tokens
  • Hybrid architecture using Gated Delta Networks and MoE
  • Unified vision-language foundation training

What it is

Qwen3.5-397B-A17B is a foundation model developed by Alibaba Cloud. It serves as the base for the hosted Qwen3.5-Plus service. The model integrates visual and textual data through early fusion training. It supports 201 languages and dialects, aiming for global deployment. With 397 billion parameters in total, it uses a sparse design to activate only 17 billion during inference. This structure seeks to balance high performance with lower computational costs.

How it came about

The model represents an advancement over the previous Qwen3 generation. Developers focused on improving multimodal learning and architectural efficiency. The release includes post-trained weights compatible with standard frameworks like Hugging Face Transformers. It builds upon earlier work in reinforcement learning and agent-based environments to enhance real-world adaptability.

Inside the model

The architecture employs a hybrid design combining Gated Delta Networks with sparse Mixture-of-Experts (MoE). Each layer routes tasks to ten active experts plus one shared expert. The model uses linear attention heads and rotary position embeddings. It supports a native context window of 262,144 tokens, extendable to over one million. Training involved multi-step prediction and scalable reinforcement learning across millions of agent environments.

What it can do

The model performs well in reasoning, coding, and visual understanding tasks. It achieves cross-generational parity with Qwen3 and outperforms Qwen3-VL variants in several benchmarks. Scores are reported for tests like HLE-Verified, MMLU-ProX, and MathVision. The system supports tool use, including code interpretation and search agents, often employing context-folding strategies to manage long interactions.

How to use it

Weights are available openly via Hugging Face for self-hosting. Users can run the model using inference engines such as vLLM or SGLang. Alibaba Cloud also offers a managed API through Model Studio, where the model is known as Qwen3.5-Plus. This hosted version includes additional production features like adaptive tool use and extended context handling by default.

Questions people ask about Qwen3.5-397B-A17B

What is the difference between Qwen3.5-397B-A17B and Qwen3.5-Plus?

Qwen3.5-397B-A17B refers to the specific model weights available for download. Qwen3.5-Plus is the hosted API version provided by Alibaba Cloud Model Studio, which includes extra production features like built-in tools and adaptive context management.

How many parameters does the model use during inference?

Although the model has 397 billion parameters in total, it uses a sparse Mixture-of-Experts architecture. This means only 17 billion parameters are activated for each token processed, reducing computational load.

Can this model process images?

Yes, it is a unified vision-language model. It undergoes early fusion training on multimodal tokens, allowing it to handle visual understanding tasks alongside text reasoning and coding.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Qwen3.5-397B-A17B yet, so the facts on this page come from the sources listed above: huggingface.co/Qwen/Qwen3.5-397B-A17B. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.