Alibaba open-sources three medium-sized Qwen 3.5 models that can be directly deployed on consumer-grade graphics cards

2026-03-09 14:34:42

Abstract generation in progress

The Beijing News Shell Finance News (Reporter Luo Yidan) — On February 25th, following the open-source release of Qwen3.5-397B-A17B on New Year’s Eve, Alibaba continued to open-source the Qianwen 3.5 series models. This release includes three medium-sized new models: Qwen3.5-35B-A3B, Qwen3.5-122B-A10B, and Qwen3.5-27B. Based on architectural innovations and training breakthroughs, all three Qianwen 3.5 models set new performance records for models of their size, surpassing the larger previous flagship models Qwen3-235B-A22B and Qwen3-VL, and outperforming GPT-5 mini on multiple leaderboards.

It is noteworthy that the new Qianwen 3.5 models can even be deployed directly on consumer-grade graphics cards, making them highly developer-friendly. Currently, the hosted model Qwen3.5-Flash based on Qwen3.5-35B-A3B is available on Alibaba Cloud Baolian, with input costs as low as 0.2 yuan per million tokens. The Qianwen 3.5 models utilize a hybrid attention mechanism, combined with innovative high-sparsity MoE architecture, and are trained on larger-scale text and visual mixed tokens. The new models achieve performance improvements with fewer total parameters and active parameters.

View Original

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.