# Qwen3
Qwen3 is a family of open-weight large language models released by Alibaba's Qwen team in April 2025, distinguished by a hybrid design that lets a single checkpoint operate in either a deliberate "thinking" mode or a fast "non-thinking" mode. The launch comprised eight checkpoints: six dense models from 0.6B to 32B parameters and two mixture-of-experts (MoE) models, topped by Qwen3-235B-A22B with 235 billion total parameters and 22 billion activated per token.[1](https://arxiv.org/pdf/2505.09388) [TechCrunch](https://www.edgepedia.org/techcrunch) reported the release on April 28, 2025, noting that the models were available for download under an open license on [Hugging Face](https://www.edgepedia.org/hugging-face) and GitHub.[2](https://techcrunch.com/2025/04/28/alibaba-unveils-qwen-3-a-family-of-hybrid-ai-reasoning-models/)
| Key fact | Value |
|---|---|
| Launch | April 2025, eight open-weight checkpoints (six dense, two MoE)[1](https://arxiv.org/pdf/2505.09388) |
| Flagship | Qwen3-235B-A22B: 235B total parameters, 22B activated[1](https://arxiv.org/pdf/2505.09388) |
| Training data | ~36 trillion tokens covering 119 languages and dialects (vendor-reported)[1](https://arxiv.org/pdf/2505.09388) |
| Context at launch | 32K tokens for 0.6B and 1.7B; 128K for 4B and larger[1](https://arxiv.org/pdf/2505.09388) |
| 2507 refresh | July 2025: separate Instruct and Thinking variants in three sizes, 256K native context extendable to 1M[3](https://github.com/qwenLM/qwen3) |
| License | Apache 2.0 for all open-weight models[3](https://github.com/qwenLM/qwen3) |
| Flagship scores (vendor) | AIME'24 85.7, AIME'25 81.5, LiveCodeBench v5 70.7, CodeForces 2,056, BFCL v3 70.8[1](https://arxiv.org/pdf/2505.09388) |
## Versions and release timeline
The April 2025 launch shipped the eight checkpoints described above as single hybrid models. In July 2025 the team released the Qwen3-2507 update, which abandoned the hybrid design in favor of two separate variants per size, Qwen3-Instruct-2507 and Qwen3-Thinking-2507, in three sizes: 235B-A22B, 30B-A3B, and 4B.[3](https://github.com/qwenLM/qwen3) The 2507 models support 256K-token native long-context understanding, extendable up to 1 million tokens; on August 8, 2025, Alibaba enabled 1M-token inputs through updated model cards.[3](https://github.com/qwenLM/qwen3)
Adjacent releases sit outside this article: Qwen3-Max, Qwen3-Next, Qwen3-[Embedding](https://www.edgepedia.org/embedding), and [Qwen3-TTS](https://www.edgepedia.org/qwen3-tts) are covered in their own articles.
## Architecture and hybrid reasoning
**Hybrid reasoning** means one model serves both deliberate and instant answering. In thinking mode, the model reasons step by step before delivering its final answer; in non-thinking mode it provides quick, near-instant responses. The mode is switchable per request within the same checkpoint.[4](https://web.archive.org/web/20251007201538/https:/qwenlm.github.io/blog/qwen3/) Mechanically, this behavior was produced by a four-stage post-training pipeline: long chain-of-thought cold start, reasoning-based reinforcement learning, thinking mode fusion, and general RL.[1](https://arxiv.org/pdf/2505.09388) The July 2025 2507 release later split these behaviors into separate Instruct and Thinking checkpoints.[3](https://github.com/qwenLM/qwen3)
The dense models use grouped-query attention (GQA), SwiGLU activations, rotary position embeddings (RoPE), and RMSNorm with pre-normalization. Compared with Qwen2, the team removed QKV-bias and added QK-Norm to the attention mechanism for training stability.[1](https://arxiv.org/pdf/2505.09388)
The MoE models have 128 total experts with 8 activated per token and, unlike Qwen2.5-MoE, exclude shared experts. Qwen3-30B-A3B has 48 layers and Qwen3-235B-A22B has 94 layers.[1](https://arxiv.org/pdf/2505.09388) Qwen documentation notes that MoE models were reintroduced with these two checkpoints, with Qwen3-32B the largest dense model.[5](https://qwen.readthedocs.io/en/latest/getting_started/concepts.html)
Pre-training ran in three stages: over 30 trillion tokens at a 4,096-token sequence length, an additional 5 trillion knowledge-intensive tokens, and finally long-context data that extended the context window to 32K.[1](https://arxiv.org/pdf/2505.09388) [Synthetic training data](https://www.edgepedia.org/synthetic-training-data) was generated with Qwen2.5-Math and Qwen2.5-Coder, and PDF text was extracted using Qwen2.5-VL.[1](https://arxiv.org/pdf/2505.09388)
## By the numbers
All Qwen3 models were pre-trained on roughly 36 trillion tokens covering 119 languages and dialects, which Alibaba describes as about twice Qwen2.5's token count and three times its language coverage.[1](https://arxiv.org/pdf/2505.09388) TechCrunch relayed the same figures with attribution to Alibaba.[2](https://techcrunch.com/2025/04/28/alibaba-unveils-qwen-3-a-family-of-hybrid-ai-reasoning-models/)
Context lengths at launch were 32K tokens for the 0.6B and 1.7B models and 128K tokens for the 4B, 8B, 14B, 32B, and both MoE models.[1](https://arxiv.org/pdf/2505.09388) The 2507 refresh raised native context to 256K, extendable to 1M tokens.[3](https://github.com/qwenLM/qwen3)
## Benchmarks: vendor claims versus independent evidence
The benchmark record for Qwen3 is almost entirely vendor-reported. According to the technical report, Qwen3-235B-A22B scored 85.7 on AIME'24, 81.5 on AIME'25, 70.7 on [LiveCodeBench](https://www.edgepedia.org/livecodebench) v5, 2,056 on CodeForces, and 70.8 on BFCL v3, surpassing [DeepSeek-R1](https://www.edgepedia.org/deepseek-r1) and [DeepSeek-V3](https://www.edgepedia.org/deepseek-v3) among open-source models, with claims of competitiveness against OpenAI-o1, Gemini-2.5-Pro, and GPT-4o.[1](https://arxiv.org/pdf/2505.09388) The report also states that Qwen3-235B-A22B-Base outperforms DeepSeek-V3 Base, Llama-4-Maverick Base, and Qwen2.5-72B-Base on most tasks with fewer total or activated parameters.[1](https://arxiv.org/pdf/2505.09388)
TechCrunch relayed these claims with attribution rather than independent verification: on [Codeforces](https://www.edgepedia.org/codeforces) the flagship "just beats OpenAI's o3-mini and Google's Gemini 2.5 Pro," and the dense Qwen3-32B surpasses OpenAI's o1 on several tests including LiveCodeBench while remaining competitive with DeepSeek's R1, per Alibaba's evaluations.[2](https://techcrunch.com/2025/04/28/alibaba-unveils-qwen-3-a-family-of-hybrid-ai-reasoning-models/) For the 2507 generation, the vendor claims Qwen3-Thinking-2507 achieves state-of-the-art results among open-weight thinking models on reasoning benchmarks.[3](https://github.com/qwenLM/qwen3)
No independent leaderboard evaluation, audit, or third-party benchmark of Qwen3 appears in the retrieved record, so the comparative scores above should be read as Alibaba's own evaluations, relayed by journalism, not as independently confirmed results.
## Licensing and availability
All open-weight Qwen3 models are licensed under Apache 2.0.[3](https://github.com/qwenLM/qwen3) At launch, all eight checkpoints (two MoE, six dense) were open-weighted, with weights published on Hugging Face, ModelScope, and Kaggle.[4](https://web.archive.org/web/20251007201538/https:/qwenlm.github.io/blog/qwen3/) The launch blog recommends SGLang and vLLM for deployment, and lists Ollama, LM Studio, MLX, llama.cpp, and KTransformers as recommended local tools.[4](https://web.archive.org/web/20251007201538/https:/qwenlm.github.io/blog/qwen3/) The retrieved record contains only these vendor recommendations, not adoption data explaining any default status on local-inference platforms.
## Adoption and reception
Alibaba's efficiency claims target small-model users: the company reported that Qwen3-30B-A3B outcompetes [QwQ-32B](https://www.edgepedia.org/qwq-32b) with 10 times fewer activated parameters, and that Qwen3-4B rivals Qwen2.5-72B-Instruct.[4](https://web.archive.org/web/20251007201538/https:/qwenlm.github.io/blog/qwen3/) These are vendor-reported comparisons; the retrieved sources do not include independent measurements of the small variants, production deployment data, or download statistics.
## References
1. [Qwen3 Technical Report (arXiv)](https://arxiv.org/pdf/2505.09388)
2. [Alibaba unveils Qwen3, a family of 'hybrid' AI reasoning models (TechCrunch, April 28, 2025)](https://techcrunch.com/2025/04/28/alibaba-unveils-qwen-3-a-family-of-hybrid-ai-reasoning-models/)
3. [QwenLM/Qwen3 (GitHub repository)](https://github.com/qwenLM/qwen3)
4. [Qwen3: Think Deeper, Act Faster (official launch blog, archived)](https://web.archive.org/web/20251007201538/https:/qwenlm.github.io/blog/qwen3/)
5. [Key Concepts - Qwen documentation](https://qwen.readthedocs.io/en/latest/getting_started/concepts.html)
---
*Topic: Encyclopedia › Technology and the built world › Computing and digital systems › Modern AI: foundation models, generative AI, and the AI industry › Model families and named models › Large language model families › Open-weight model families*
*Initially written Sep 17, 2026 · Edited: Oct 11, 2026; Oct 15, 2026*
*Copyright 2026 EdgeChat AI, a subsidiary of Biostate AI.*
License: Edgepedia Community License 1.0, https://www.edgepedia.org/edgepedia/license