qwen3.5-397b-a17b

qwen3.5-397b-a17b

Alibaba’s Qwen3.5 series supports multimodal capabilities.
2026-02-16
LLM
Model capability: imageModel capability: videoModel capability: thinkingModel capability: function_call
Input:
$0.171/1M tokensstarting from
Output:
$1.03/1M tokensstarting from
Bulk order? Contact your manager for exclusive deals

API Overview

Qwen3.5-397B-A17B is a flagship foundation model released by Tongyi Qianwen, boasting 397 billion parameters (17 billion activated parameters). Built on a brand-new Hybrid Mixture-of-Experts architecture, this model deeply integrates multimodal capabilities with underlying logical reasoning. As the performance cornerstone of the Qwen3.5 series, it not only demonstrates cross-generational-leading advantages in benchmark tests but also achieves technological breakthroughs in inference efficiency, multilingual coverage, and the execution of complex agent tasks, providing developers with an underlying paradigm that combines deep understanding and ultimate performance. ───────────────────────────────────────────────────────────────────

Core Capabilities


Efficient Mixture-of-Experts (MoE) Architecture: Employing Gated Delta Networks and sparse MoE technology, this architecture maintains an ultra-large knowledge base while significantly reducing inference latency and computational costs, striking a perfect balance between high performance and high throughput. Native Multimodal Fusion: Adopting an Early Fusion training strategy, it enables visual and textual features to be aligned at the very bottom layer, greatly enhancing the model's performance in visual question answering, complex document recognition, and cross-modal logical reasoning. Ultimate Inference and Scalability: Featuring a native context window of 260,000 tokens and supporting technical extensions beyond 1 million tokens, this model boasts powerful chain-of-thought generation capabilities, enabling it to tackle extremely challenging tasks such as mathematical competitions, hardcore programming, and complex planning. Global Language and Environmental Adaptability: Covering 201 languages and dialects, combined with reinforcement learning training in large-scale agent environments, this model exhibits exceptional cross-cultural understanding and real-world task adaptability.

Playground

Log in to explore more features! Click to Log In

API Analytics

API Reference (1)

API DescriptionAPI EndpointRequest MethodStabilityParameter Description
qwen3.5-397b-a17b
POST
Stable
View Details

API Pricing

$
ModelDescriptionContextOfficial Price302.AI Price

qwen3.5-397b-a17b

Input <= 128K
254000

Input$0.171 / 1M tokens
Output$1.03 / 1M tokens

Input$0.171/ 1M tokens
Output$1.03/ 1M tokens
Original Price

qwen3.5-397b-a17b

128K-256K
254000

Input$0.43 / 1M tokens
Output$2.6 / 1M tokens

Input$0.43/ 1M tokens
Output$2.6/ 1M tokens
Original Price