MDL-2109 EST.2025 · IDX.569 Language model In production ⇄ Compare
Qwen 3 Alibaba Cloud · China · 2025
The world's most-downloaded open model family — 100+ languages, from 0.6B to 235B parameters.
Alibaba's Qwen 3 spans tiny edge models to a 235B-parameter MoE flagship, with hybrid thinking modes that toggle between fast answers and deep reasoning. Supporting 100+ languages under an Apache-2.0 license, Qwen became the most-downloaded open model family on Hugging Face and the backbone of countless fine-tunes worldwide.
#open weights #multilingual #MoE #most-downloaded
Official website ↗ ▸ SpecificationsClass & identity
Model class Open-weights LLM family
Version / release 2025 Architecture
Architecture Dense + Mixture-of-Experts (hybrid thinking)
Parameters 0.6B – 235B (MoE flagship)
Context window Up to 128k+ tokens
Modalities Text (+ multimodal variants)
Knowledge cutoff / data — Performance (benchmarks)
MMLU —
GPQA Diamond —
SWE-bench —
MATH / AIME —
Chatbot Arena Elo —
Domain benchmark — Capabilities Text Extended reasoning Open weights Fine-tunable On-device Multilingual Agentic coding Function calling
Reasoning mode Hybrid thinking modes (toggle)
Tool use / agentic Function calling, agentic
Multilinguality 100+ languages
Key strengths Most-downloaded open family; broad size range Access & deployment
Availability Open weights
License Apache 2.0
Pricing Free to run yourself: weights on Hugging Face under Apache 2.0, commercial use included. Hosted access is sold by third parties and by Alibaba Cloud, each at its own rates — the current Model Studio catalogue lists the newer qwen3.7 and qwen3.8 families.
Deployment Self-host / API / on-device (small sizes)
Safety / alignment Open model Compiled to the wujec.ai Spec Standard v1 — “—” means not publicly disclosed.