Qwen3
Alibaba's 2025 Apache 2.0 family from 0.6B to 235B with a switchable thinking mode. Legacy, yet still hugely downloaded.
In short. Qwen3 is Alibaba's April 2025 family of text models, from a 0.6B model small enough for a phone up to a 235B giant. One model can either think step by step or answer straight away. Newer Qwen generations have replaced it, but its small and mid-size models are still some of the most downloaded language models on Hugging Face, and they turn up in countless tutorials, tools and fine-tunes.
Summary
Alibaba's Qwen team released Qwen3 on 29 April 2025 under the Apache 2.0 license, as eight open models: six dense models (0.6B, 1.7B, 4B, 8B, 14B and 32B) and two mixture-of-experts models (30B-A3B and the flagship 235B-A22B with 22B active parameters). Its signature feature is a hybrid design in which one model supports a thinking mode that reasons step by step and a non-thinking mode that answers quickly, switched through the chat template, with a thinking budget to trade quality against latency. Qwen says pre-training used about 36 trillion tokens covering 119 languages and dialects, nearly twice Qwen2.5's 18 trillion. The native context is 32,768 tokens, extendable to 131,072 with YaRN, and the technical report is arXiv 2505.09388.
In July and August 2025 Qwen split the two modes into separate "2507" releases, such as Qwen3-30B-A3B-Instruct-2507, Qwen3-235B-A22B-Instruct-2507 and Qwen3-4B Instruct and Thinking, with a native context of 262,144 tokens. In the same weeks it released Qwen3-Coder, agentic coding models including Qwen3-Coder-480B-A35B and Qwen3-Coder-30B-A3B, also with a 262,144-token context. We cover all of these under this one legacy entry.
We mark Qwen3 as legacy because Qwen3.5 and later generations replace it, and those newer models also take images and video natively. Usage is still very high: Hugging Face showed about 20.2 million downloads for Qwen3-0.6B, 12.9 million for Qwen3-14B and 6.4 million for Qwen3-8B in the 30 days to 11 November 2026, because they are default test models, fine-tuning bases and small local models. For a new project pick a newer Qwen unless a tool or fine-tuning recipe is tied to Qwen3.
Good for
- A well-supported small or mid-size model for local use and experiments (0.6B to 14B), with 4-bit builds in nearly every runner
- Fine-tuning and distillation experiments, since so many recipes and adapters target Qwen3
- Switching between fast answers and step-by-step reasoning in a single model
- Multilingual use across 119 languages and dialects
Watch out for
- Marked legacy because Qwen3.5 and later models are newer and take images and video natively
- The April models have a native context of only 32K (131K with YaRN), far shorter than recent 256K to 1M models, and only the 2507 updates raise it to 262,144
- The April models combine both modes, while the 2507 updates split them into separate instruct and thinking models, so check which one you download
- The benchmark claims in Qwen's blog are self-reported
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the 0.6B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| 0.6B BF16 | 752M | 1.73 GB | 0.865 GB | 0.519 GB | Laptop or phone |
| 1.7B BF16 | 2.03B | 4.67 GB | 2.34 GB | 1.4 GB | Laptop or phone |
| 4B BF16 | 4.02B | 9.25 GB | 4.63 GB | 2.78 GB | Laptop or phone |
| 8B BF16 | 8.19B | 19 GB | 9.42 GB | 5.65 GB | Laptop or phone |
| 14B BF16 | 14.8B | 34 GB | 17 GB | 10 GB | Laptop or phone |
| 30B-A3B BF16 | 30.5B 3B active | 70 GB | 35 GB | 21 GB | Gaming GPU |
| 32B BF16 | 32.8B | 75 GB | 38 GB | 23 GB | Gaming GPU |
| 235B-A22B BF16 | 235B 22B active | 541 GB | 270 GB | 162 GB | Big-memory desktop |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
Frequently asked questions
- What is Qwen3?
- Qwen3 is Alibaba's April 2025 family of text models, from a 0.6B model small enough for a phone up to a 235B giant. One model can either think step by step or answer straight away. Newer Qwen generations have replaced it, but its small and mid-size models are still some of the most downloaded language models on Hugging Face, and they turn up in countless tutorials, tools and fine-tunes.
- Can I use Qwen3 commercially?
- Yes. Qwen3 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Qwen3 open source?
- Qwen3 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Qwen3 on my own computer?
- Yes. The smallest variant, 0.6B, needs about 0.519 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Qwen3 released?
- Qwen3 was released by Alibaba Qwen on April 29, 2025.
- How long a context window does Qwen3 have?
- Qwen3 supports a context window of up to 32K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Qwen blog: Qwen3, Think Deeper, Act Faster (29 Apr 2025) qwenlm.github.io
- Qwen3-8B model card (Hugging Face) huggingface.co
- Qwen3-4B-Instruct-2507 model card (Hugging Face) huggingface.co
- Qwen3-Coder-30B-A3B-Instruct model card (Hugging Face) huggingface.co
- Qwen3 technical report (arXiv 2505.09388) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3.5 Alibaba Qwen | 873M – 403B 17B active | 256K | Apache 2.0 (Permissive) | 0.602 GB Laptop or phone | |
| Qwen3.6 Alibaba Qwen | 27.8B – 36B 3B active | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Qwen3-VL Alibaba Qwen | 2.13B – 236B 22B active | 256K | Apache 2.0 (Permissive) | 1.47 GB Laptop or phone | |
| gpt-oss OpenAI | 21B – 117B 5.1B active | 128K | Apache 2.0 (Permissive) | 15 GB Laptop or phone |