Home / Technology / Alibaba AI Shatters Performance Barriers
Alibaba AI Shatters Performance Barriers
26 Feb
Summary
- New open-source AI models rival proprietary West-coast giants.
- Models achieve over 1 million token context length on consumer GPUs.
- Hybrid architecture integrates Gated Delta Networks and MoE.
Alibaba's Qwen AI team has launched the Qwen 3.5 Medium Model series, featuring four new large language models. Three of these models are available under the Apache 2.0 license for commercial use, accessible via Hugging Face and ModelScope. These open-source models demonstrate performance comparable to, and in some cases exceeding, proprietary models from OpenAI and Anthropic, according to third-party benchmarks.
A significant advancement is the Qwen 3.5 series' ability to handle context windows exceeding 1 million tokens on consumer-grade GPUs with 32GB of VRAM. This is facilitated by near-lossless accuracy under 4-bit weight and KV cache quantization. The models utilize a hybrid architecture combining Gated Delta Networks with a sparse Mixture-of-Experts (MoE) system.
These new models feature a default "Thinking Mode," where they generate an internal reasoning chain before providing a final answer. Alibaba Cloud Model Studio also offers Qwen3.5-Flash via API, noted for its cost-effectiveness compared to Western models, with specific pricing for features like Web Search and Code Interpreter.
The Qwen 3.5 release enables advanced AI development for on-premise use, allowing organizations to process large datasets locally without relying on server-grade infrastructure. This enhances data security and control, enabling deep institutional analysis without privacy risks associated with third-party APIs.
