Alibaba published the Qwen3 Technical Report, introducing dense (0.6B–32B) and MoE (up to 235B total / 22B active) models trained on 36 trillion tokens with a unified thinking-budget switch inside a single checkpoint.
Qwen
Also known as: Qwen2.5 · Qwen3 · Qwen-VL · Alibaba Qwen
Plain English
Alibaba's flagship family of open-weight language, coding, and multimodal models, spanning tiny laptop-sized checkpoints up to frontier-scale Mixture-of-Experts models.
Technical Overview
Developed by Alibaba Cloud's Qwen team, the Qwen lineage spans dense and sparse MoE Transformers alongside vision-language (Qwen-VL) and coding (Qwen-Coder) variants under permissive Apache-2.0 weights. Qwen3 unified hybrid thinking (controllable chain-of-thought) and non-thinking modes inside single checkpoints across 0.6B–235B parameters, making Qwen one of the most widely fine-tuned and distilled open-weight bases.
Chronological Entries (3)
The Qwen team published the Qwen2.5 Technical Report covering models from 0.5B to 72B parameters trained on 18 trillion tokens. Qwen2.5-Coder and Qwen2.5-Math became dominant open-weight foundations for Post-Training (SFT, RLHF, DPO & GRPO), Distillation & Synthetic Data, and local GGUF serving.
Alibaba published the Qwen Technical Report, open-sourcing base and chat Checkpoints alongside specialized coding and math variants trained on trillions of multilingual tokens—establishing the Qwen lineage as a core open-weight alternative to Llama.
Direct Relations (5)
- Created byChinese AI Labs
- ImplementsTransformer
- ImplementsMixture of Experts (MoE)
- Alternative toLlama
- Alternative toDeepSeek
Referenced By (0)
No incoming references yet.