Qwen3 (235B MoE)
text
open-source
Qwen3 is Alibaba's open-source LLM family. Its largest 235B-model (22B active) supports 119 languages. Trained on 36T to...
Version: 235B
Released: 6m 3d ago on 04/29/2025
Pricing:
- details: free
Repository: Hugging Face (Alibaba/Qwen3)
Architecture
- parameters: 235 billion (22B active)
- context_length: 38,000 (max)
- inference_type: Hybrid MoE Transformer (thinking vs. non-thinking)
- training_data: 36 trillion tokens multilingual
Capabilities
- 119 languages
- multi-language support
- multilingual reasoning
- code generation
- translation
- agentic tool use
Benchmarks
- Math/Code: Top-tier on AIME25 (math) and LiveCodeBench (coding)
Safety
- Aligned with Alibaba's policies (improved for user-friendly, creative outputs)
- Reinforcement learning and chain-of-thought tuning to enhance reliability
- Biases may appear in less common languages
Deployment
- regions: Global (China-focused)
- hosting: GitHub, ModelScope, Alibaba Model Studio (API)
- integrations: Used in Alibaba's Quark AI assistant
Tags
LLMopen-sourceMoEmultilingual