Qwen3

// Overview

Qwen3 is the third-generation open-weight large language model family from Alibaba's Qwen team, released on April 28, 2025. Rather than a single model, it is a full lineup: dense models at 0.6B, 1.7B, 4B, 8B, 14B, and 32B parameters, plus two Mixture-of-Experts models — Qwen3-30B-A3B (30B total, 3B active) and the flagship Qwen3-235B-A22B (235B total, 22B active). Every model is released under the permissive Apache 2.0 license with downloadable weights on Hugging Face, ModelScope, Kaggle, and Ollama.

Qwen3's headline feature is a hybrid design: a single model can operate in a thinking mode that produces a step-by-step reasoning trace before its final answer, or a non-thinking mode that responds directly for low-latency chat. Users (and developers, via a flag or a /think and /no_think control) choose how much the model deliberates per task, so one deployment covers both hard reasoning and fast dialogue without swapping models. Qwen3 was pre-trained on roughly 36 trillion tokens — nearly double Qwen2.5 — spanning 119 languages and dialects, and post-trained with reinforcement learning for math, code, and agentic tool use.

The flagship Qwen3-235B-A22B activates only 22B of its 235B parameters per token, giving it inference cost closer to a mid-size dense model while competing with much larger systems on reasoning and coding benchmarks. The smaller, edge-friendly dense models (0.6B-8B) make Qwen3 practical to run locally on a laptop or single GPU. Hosted API access is available through Alibaba Cloud Model Studio and aggregators such as OpenRouter.

// Benchmarks

1. AIME 2024 (math, thinking mode)85.7%
2. AIME 2025 (math, thinking mode)81.5%
3. LiveCodeBench v5 (coding)70.7%
4. Arena-Hard (alignment)95.6%
5. BFCL v3 (function calling)70.8%
6. MMLU-Pro (knowledge)68.18%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Rates shown for the flagship Qwen3-235B-A22B via OpenRouter (a 35%-off promotional rate at time of capture). The open weights are free to self-host; smaller dense models are far cheaper or free to run locally.

Pricing source ↗

// Strengths

• Fully open weights under Apache 2.0 — commercial-friendly, self-hostable, and downloadable from Hugging Face, ModelScope, Kaggle, and Ollama
• One model, two modes: switch between a deliberate thinking mode and a fast non-thinking mode without changing models
• Wide size range, from a 0.6B dense model that runs on-device to a 235B-A22B MoE flagship for frontier-level reasoning
• Efficient MoE flagship — only 22B of 235B parameters activate per token, keeping inference cost low relative to total scale
• Strong math and coding results (AIME, LiveCodeBench) and high Arena-Hard alignment scores among open models
• Broad multilingual coverage: pre-trained across 119 languages and dialects

// Best for

• Self-hosted or private-cloud assistants where open weights and Apache 2.0 licensing are required
• On-device and edge deployment using the small dense models (0.6B-8B) on laptops or single GPUs
• Math, coding, and STEM problem-solving that benefits from the thinking (reasoning) mode
• Cost-sensitive, high-volume chat served in non-thinking mode for low latency
• Agentic and tool-using workflows that call functions across multiple turns
• Multilingual applications that need coverage well beyond English and Chinese

// How to access

// Qwen (open-weight) — every version

The full lineage of the Qwen (open-weight) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.