Alibaba Qwen

Latest AI news, models and releases from Alibaba Qwen. ['Qwen', 'Qwen2.5', 'Qwen2.5-14B', 'Qwen2.5-32B', 'Qwen2.5-3B', 'Qwen2.5-72B', 'Qwen2.5-7B', 'Qwen2.5-Coder', 'Qwen2.5-Math', 'Qwen3-0.6B', 'Qwen3-14B', 'Qwen3-1.7B', 'Qwen3-235B-A22B', 'Qwen3-30B-A3B', 'Qwen3-30B-A3B-Base', 'Qwen3-32B', 'Qwen3-4B', 'Qwen3-8B', 'Qwen3 Coder', 'Qwen3 Instruct', 'Qwen3 Thinking', 'Qwen-Plus', 'Qwen-Turbo', 'QwQ-32B']

Models 🇨🇳

Alibaba Qwen Releases Qwen3: Hybrid Thinking, Support for 119 Languages, and Competition with Leaders

Alibaba Qwen has announced the Qwen3 model family, including the flagship Qwen3-235B-A22B with 235 billion parameters (22 billion active), which is compared to DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro. The models support hybrid thinking modes (reasoning and fast response), work with 119 languages, feature improved agent capabilities and MCP integration. Weights of eight models are released under the Apache 2.0 license.

Alibaba/Qwen

Alibaba Qwen28.07 · 00:03

Research 🇨🇳

GSPO: scalable reinforcement learning for language models

Alibaba Qwen introduced the Group Sequence Policy Optimization (GSPO) algorithm to address instability and model collapse during long RL training. GSPO uses sequence-level optimization, offering higher efficiency and robustness than GRPO, especially for MoE models. The algorithm underpins the Qwen3 models.

Alibaba/Qwen

Alibaba Qwen27.07 · 17:05

Show more ↓