Qwen 3
Alibaba's open-weight model family with switchable thinking and non-thinking modes.
Context window
128K tokens (32K for 0.6B/1.7B/4B dense variants)
Input / 1M tokens
Free
Output / 1M tokens
Free
Provider
Alibaba (Qwen Team)
Open-weight models under Apache 2.0 - free to download and self-host. Running costs depend on your hardware or inference provider pricing (e.g., Together AI, Fireworks, Ollama, Hugging Face). Alibaba Cloud Model Studio also offers hosted API access. · Data verified 2026-07-02
Qwen3 is Alibaba's open-weight large language model series, released April 29, 2025, under Apache 2.0. It spans two MoE models - Qwen3-235B-A22B (235B total / 22B active) and Qwen3-30B-A3B (30B total / 3B active) - plus six dense models (0.6B, 1.7B, 4B, 8B, 14B, 32B). Its signature feature is a dual-mode design: a 'thinking' mode for step-by-step reasoning and a 'non-thinking' mode for fast responses, controllable per task. Larger models support a 128K-token context window (32K on the smallest dense variants), and the series supports 119 languages and dialects with strong coding, math, and agentic performance.
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Download open weights from Hugging Face, ModelScope, or Kaggle and run via Transformers, vLLM, SGLang, or Ollama. Also available through Alibaba Cloud Model Studio and third-party inference providers.
Strengths
- ✓Open weights under permissive Apache 2.0 license
- ✓Switchable thinking / non-thinking modes for quality-vs-speed control
- ✓Wide range of sizes from 0.6B to 235B (MoE)
- ✓Efficient MoE variants (e.g., 30B-A3B activates only 3B parameters)
- ✓Broad multilingual support (119 languages and dialects)
Best for developers who...
When to choose it (and when not to)
Reach for Qwen 3 when...
- →You need an open, self-hostable model with a permissive license
- →You want to toggle deep reasoning on or off per request
- →Multilingual applications
- →Efficient inference via MoE with few active parameters
Look elsewhere if...
- ✕You want a fully managed API with no infrastructure (though hosted options exist)
- ✕You need native multimodal input (base Qwen3 dense/MoE releases are text-focused)
- ✕You require the newest frontier quality (newer Qwen 3.x releases and frontier hosted models exist)
How to use it
- ›Enable thinking mode for hard reasoning/math/coding; disable it for fast, cheap responses
- ›Use the instruction-tuned checkpoints and Qwen chat template
- ›Pick an MoE variant (30B-A3B) for a strong quality-to-compute ratio
- ›Quantize to fit your GPU; smaller dense models suit edge deployment
Quickstart
Pythonfrom transformers import pipeline
pipe = pipeline("text-generation", model="Qwen/Qwen3-235B-A22B", device_map="auto")
messages = [{"role": "user", "content": "Solve: what is the derivative of x^3?"}]
print(pipe(messages, max_new_tokens=256)[0]["generated_text"])Install `transformers` and `accelerate`. For a smaller footprint use Qwen/Qwen3-30B-A3B or Qwen/Qwen3-8B, or run locally with `ollama run qwen3`.
API model id: Qwen/Qwen3-235B-A22B
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| Qwen3-235B-A22B | 235B total / 22B active | Flagship MoE; competitive on coding, math, and general benchmarks |
| Qwen3-30B-A3B | 30B total / 3B active | Outcompetes QwQ-32B with ~10x fewer active parameters per Qwen |
Source: Qwen team - Qwen3 blog
Compare Qwen 3 with
Qwen 3 vs DeepSeek V3
DeepSeek - 128K tokens ctx
Qwen 3 vs Llama 4
Meta - Up to 10M tokens (Scout); ~1M tokens (Maverick) ctx
Qwen 3 vs Gemma 3
Google DeepMind - 128K tokens (32K for the 1B variant) ctx
Qwen 3 vs Mistral Large
Mistral AI - 128000 ctx