Qwen 3: The Complete Guide

Qwen 3 is Alibaba Cloud's most ambitious open-source AI initiative to date. Launched in April 2025 and continuously expanded ever since, the Qwen 3 family now spans text LLMs, coding agents, vision-language models, speech recognition, voice synthesis, embeddings, and a trillion-parameter cloud flagship. Every open-weight variant ships under the permissive Apache 2.0 license, trained on 36 trillion tokens across 119 languages — and competes head-to-head with GPT-5, Gemini 3, and Claude Opus. Explore the full lineup on the Qwen AI homepage.

From a 0.6B-parameter edge model that fits on a Raspberry Pi to the 1-trillion-parameter Qwen3-Max-Thinking that tops the HLE leaderboard, the Qwen 3 ecosystem covers every deployment scenario. Stand-out features include a Hybrid Reasoning Engine, sparsely-activated MoE architectures, context windows up to 1 million tokens, and native tool calling. Below you'll find the complete model catalog, benchmarks, local deployment guides, and API pricing.

The Qwen 3 Ecosystem

What started as eight text models in April 2025 has evolved into a full-stack AI platform. Here's every sub-family at a glance:

This page focuses on the core text LLMs (base models, 2507 update, and Qwen3-Max). For specialized models, follow the links above to their dedicated guides.

Base LLM Line-up: Dense and MoE

The original April 2025 launch introduced six dense models (0.6B to 32B) and two Mixture-of-Experts variants (30B-A3B and 235B-A22B). All share the same tokenizer, instruction format, and hybrid thinking/non-thinking capability — meaning you can swap model sizes without rewriting your code.

Dense Models (0.6B → 32B)

Ideal for chatbots, real-time RAG pipelines, and edge deployment. The 8B variant runs comfortably under 10 GB VRAM with GGUF quantization while matching much larger competitors on multilingual benchmarks.

Mixture-of-Experts (MoE) Variants

• Qwen3-30B-A3B — 30.5B total parameters, ~3.3B active per token. Runs in real time on a single RTX 4090 and delivers Arena-Hard scores of 91.0 — at 1/8th the GPU cost of comparable dense models.
• Qwen3-235B-A22B — 235B total, ~22B active (8 of 128 experts per token). The open-source flagship that rivals GPT-4-class reasoning while halving inference cost compared to a monolithic dense model of equivalent quality.

Specification Matrix (April 2025 Launch)

The Qwen3-2507 Update (July 2025)

Three months after the initial launch, the Qwen team released a major revision that changed the reasoning architecture philosophy. Instead of a single hybrid model that switches between thinking and non-thinking modes, the 2507 update introduced dedicated variants:

• Instruct-2507 — Non-thinking only. Optimized for instruction following, chat, and tool use. No <think> blocks generated.
• Thinking-2507 — Thinking only. Always reasons through problems. Optimized for math, STEM, and complex logic.

2507 Variant Matrix

Key Improvements in 2507

• Context window jump: From 32K native / 128K extended to 256K native and up to ~1 million tokens using DCA + MInference sparse attention. The 235B-Instruct-2507 scores 82.5% accuracy on the RULER benchmark at 1M tokens.
• Massive non-thinking gains: The Instruct-2507 variants saw dramatic improvements — for example, AIME25 jumped from 24.7 to 70.3 and ZebraLogic from 37.7 to 95.0 on the 235B model.
• Better instruction following: Enhanced math, coding, tool use, and multilingual knowledge across all sizes.
• The 4B-Thinking-2507 achieves AIME25 scores of 81.3 — rivaling the much larger Qwen2.5-72B-Instruct.

Qwen3-Max & Qwen3-Max-Thinking

Qwen3-Max is Alibaba's closed-source flagship — a 1+ trillion parameter MoE model available exclusively through API. Launched in September 2025, it received a major upgrade in January 2026 with the addition of test-time scaling (TTS) under the name Qwen3-Max-Thinking.

The TTS mechanism isn't naive best-of-N sampling — it uses an experience-cumulative, multi-round reasoning strategy that progressively refines answers across multiple internal passes. This pushes several benchmarks to state-of-the-art levels.

What Qwen3-Max-Thinking Leads On

Qwen3-Max Specifications

Benchmark Comparison: Qwen3-Max-Thinking vs Frontier Models

Architecture Deep-Dive

All Qwen 3 text models share a Transformer backbone with several key innovations:

• Grouped Query Attention (GQA) — Reduces KV-cache memory for faster inference and longer contexts.
• QK-Norm — Stabilizes attention logits during training, especially at large scale.
• SwiGLU Activations — Replaces standard FFN layers for better parameter efficiency.
• RoPE Scaling — Rotary position embeddings extended with YaRN for context beyond the training length.

The MoE variants use a 128-expert grid (8 active per token on the 235B model) with a global-batch load-balancing loss to prevent expert collapse. This sparse routing means only ~22B parameters fire on each forward pass, cutting FLOPs by ~6× compared to a dense 235B model.

The later Qwen3-Next (September 2025) introduced a hybrid attention mechanism combining linear attention with standard self-attention for even greater efficiency, signaling the direction of future Qwen 3 releases.

The Hybrid Reasoning Engine

One of Qwen 3's signature innovations is its dual-mode reasoning system — the ability to switch between deep chain-of-thought reasoning and instant direct responses:

• Thinking Mode — The model emits explicit <think> chains before answering, walking through problems step-by-step. Ideal for math, STEM, multi-hop logic, and coding.
• Non-Thinking (Fast) Mode — Bypasses chain-of-thought entirely for low-latency responses. Best for support chat, search, simple QA, and high-throughput scenarios.
• Thinking Budget — You can hard-cap reasoning tokens with max_thought_tokens to control the latency/accuracy trade-off.

In the original April 2025 release, both modes lived in a single hybrid model controlled by the enable_thinking parameter. The 2507 update split this into dedicated Instruct (non-thinking) and Thinking variants — each optimized independently for its mode, resulting in significantly better performance on both sides.

Training & Alignment

The Qwen 3 training pipeline has three main pre-training stages followed by multi-step post-training:

Pre-training

• Stage 1 (30T tokens): Mixed-quality web and code data at 4K sequence length — builds broad world knowledge.
• Stage 2 (5T tokens): High-quality STEM, coding, and reasoning data — sharpens analytic ability.
• Stage 3 (hundreds of B tokens): 32K sequence long-context bootstrapping — teaches the model to handle extended documents.

Post-training

• Multi-round supervised fine-tuning (SFT) with diverse instruction data
• Reward-model RLHF for alignment with human preferences
• Reasoning-focused reinforcement learning to improve chain-of-thought quality
• On-policy distillation that reduces compute by ~90% compared to pure RL

Benchmarks & Performance

Here's how the key Qwen 3 open-source models perform across major benchmarks, compared to competitors:

Qwen3-235B-A22B-Thinking-2507 (Open-Source Flagship)

April 2025 Launch Benchmarks (Original Models)

Run Qwen 3 Locally

Every open-weight Qwen 3 model is available on Hugging Face and ModelScope in multiple formats. Here are the main deployment options:

Ollama (Easiest)

That's it — Ollama automatically downloads the quantized model and starts an interactive chat. Other popular tags:

• ollama run qwen3:4b — Lightweight, runs on 8GB VRAM
• ollama run qwen3:14b — Sweet spot for most users
• ollama run qwen3:32b — Near-frontier on a single GPU
• ollama run qwen3:30b-a3b — MoE: 30B params, only 3B active

vLLM (Production Serving)

vLLM provides OpenAI-compatible API endpoints with PagedAttention for maximum throughput. Recommended for multi-user deployments.

llama.cpp (CPU + GPU Hybrid)

GGUF quantized models run with partial GPU offloading, making Qwen 3 accessible even on machines with limited VRAM.

Transformers (Python)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")

messages = [{"role": "user", "content": "Explain quantum entanglement simply."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(output[0], skip_special_tokens=True))

API Access & Pricing

Alibaba Cloud offers Qwen 3 models through DashScope / Model Studio with OpenAI-compatible endpoints. The flagship Qwen3-Max-Thinking pricing:

The open-source models can also be served through third-party providers like OpenRouter, Novita AI, and Fireworks AI — often at lower per-token prices than self-hosting.

Base URL (International): https://dashscope-intl.aliyuncs.com/compatible-mode/v1

Hardware Requirements

Tip: GGUF quantization (Q4_K_M) reduces VRAM usage by 70–80% with minimal quality loss. For the 235B MoE, AWQ and GPTQ quantized builds on Hugging Face make single-node deployment feasible on 2× A100 80GB.

Fine-Tuning Qwen 3

Thanks to the Apache 2.0 license, you can fine-tune any open-weight Qwen 3 model on your own data. Common approaches:

• LoRA / QLoRA — Low-rank adaptation with 4-bit quantization. Fine-tune the 8B model on a single RTX 4090 with 24GB VRAM. Tools: Unsloth, LLaMA-Factory, or Hugging Face PEFT.
• Full fine-tuning — For maximum control. Requires multi-GPU setups for models above 8B. DeepSpeed ZeRO-3 or FSDP recommended.
• DPO / RLHF — Preference-based alignment to customize model behavior. Supported natively by TRL (Transformer Reinforcement Learning) library.

Recommended starting point: QLoRA on Qwen3-8B or Qwen3-14B with Unsloth for 2–4× speedup and 60% less memory.

Use Cases

Limitations & Considerations

While Qwen 3 is impressive, it's important to be aware of its boundaries:

• Hallucination: Like all LLMs, Qwen 3 can generate plausible-sounding but incorrect information. Always verify factual claims in critical applications.
• SWE-Bench gap: On real-world software engineering tasks (SWE-Bench Verified), Qwen3-Max-Thinking scores 75.3 — behind Claude Opus 4.5 (80.9) and GPT-5.2 (80.0).
• Speed vs. intelligence trade-off: Qwen3-Max-Thinking runs at ~38 tokens/second — significantly slower than lighter models. For latency-sensitive applications, consider the 30B-A3B or dense variants.
• Closed-source flagship: Qwen3-Max is API-only and proprietary. The most powerful Qwen model cannot be self-hosted or fine-tuned.
• MoE memory overhead: While MoE models activate fewer parameters, they still require loading all expert weights into memory. The 235B MoE needs ~80+ GB VRAM even quantized.
• Safety alignment: As with any open-weight model, fine-tuned variants may lose safety guardrails. Deploy responsibly with appropriate safeguards.

Frequently Asked Questions

Is Qwen 3 free to use commercially?

Yes — all open-weight models (0.6B through 235B) are released under the Apache 2.0 license, which permits unrestricted commercial use, modification, and distribution. Only Qwen3-Max is proprietary (API-only).

What's the best Qwen 3 model for my hardware?

For 8 GB VRAM: Qwen3-4B or Qwen3-8B (Q4 quantized). For 24 GB VRAM: Qwen3-32B or Qwen3-30B-A3B (MoE). For cloud/multi-GPU: Qwen3-235B-A22B-Thinking-2507.

Should I use the original models or the 2507 update?

If a 2507 variant exists for your target size (4B, 30B-A3B, or 235B-A22B), always prefer the 2507 version. It offers dramatically better performance, larger context windows, and optimized reasoning. For sizes without a 2507 update (0.6B, 1.7B, 8B, 14B, 32B), the original models remain the latest available.

How does Qwen 3 compare to Llama, GPT, and Gemini?

The open-source Qwen3-235B-Thinking-2507 competes directly with O4-mini and Gemini 2.5 Pro on math and coding benchmarks. The closed-source Qwen3-Max-Thinking trades blows with GPT-5.2 and Gemini 3 Pro, leading on several benchmarks (HLE, IMO, LiveCodeBench) while trailing on others (SWE-Bench).

Can I use Qwen 3 for tool calling and agents?

Yes — Qwen 3 supports native JSON-defined tool calling compatible with the Model Context Protocol (MCP). Define your tools as JSON schemas and the model will generate structured function calls automatically. No custom parsing needed.

{
  "name": "lookupWeather",
  "description": "Get current weather for a city",
  "parameters": {
    "type": "object",
    "properties": { "city": { "type": "string" } },
    "required": ["city"]
  }
}

Conclusion: The Most Complete Open AI Ecosystem

Qwen 3 isn't just a language model — it's a complete AI platform. From edge-ready 0.6B models to the trillion-parameter Qwen3-Max-Thinking that leads frontier benchmarks, from agentic coding with Qwen3-Coder-Next to voice cloning with Qwen3-TTS and 52-language transcription with Qwen3-ASR — Alibaba Cloud has built an ecosystem where every component is either open-source and free to fine-tune, or competitively priced via API.

Whether you're building an autonomous agent, deploying a multilingual support bot, running a local code assistant, or pushing the limits of mathematical reasoning, Qwen 3 has a model for your use case. Fork it on GitHub, download weights from Hugging Face, or start chatting right now at chat.qwen.ai. Explore the full ecosystem on the Qwen AI homepage.