Alibaba points to official technical details, model weights, and GitHub/social channels for updates.

The article credits Ayesha Aayat and notes a focus on cybercrime and digital safety, with external link references included.

Alibaba unveils Qwen3.5, a multimodal mixture-of-experts AI that Alibaba says outperforms GPT-5.2 and Claude 4.5 Opus on several benchmarks, while also delivering 8x better throughput on large workloads and 60% lower operating costs.

Qwen3.5 is built with native multimodality, including text, images, and video, and introduces agentic capabilities that can autonomously take actions across mobile and desktop apps.

The model scales via a sparse Mixture-of-Experts routing across 60 layers, featuring a hybrid Efficient Hybrid Architecture with Gated Delta Networks and both linear and quadratic attention components to reduce memory and compute.

Key architectural highlights include a large-context capability with a base context window of up to 256k tokens and Qwen3.5-Plus extending to 1,000,000 tokens for long inputs, enabled by an asynchronous RL framework.

Qwen3.5 supports over 210 languages and can process images and data visualizations, enabling rich multimodal input.

The available excerpt does not disclose the exact release date, deployment scope, or additional technical specs.

The architecture employs 512 MoE experts with 11 active per token (10 routing experts plus 1 shared) and a 4,096 hidden size, using a 60-layer configuration with a 3:1 ratio of gating attention to gating delta networks.

Analysts warn that rising costs, data governance, regulatory changes, and heavy computing demands could affect long-term profitability and sustainability of large models.

MoE details include 512 experts with 11 active per token and a vocabulary of around 248k tokens.

Qwen3.5’s base length is 256k tokens, with hosted Qwen3.5-Plus supporting 1,000,000 tokens to handle long documents and large codebases without full RAG.