Qwen3 Unveiled: 8 Open‑Source Hybrid Inference Models Redefine AI Capabilities

Model Release

Qwen3 is released as eight fully open‑source hybrid inference models:

Mixture‑of‑Experts (MoE) models :

Qwen3‑235B‑A22B – ~235 billion total parameters, ~220 billion activation parameters

Qwen3‑30B‑A3B – 30 billion total parameters, 30 billion activation parameters

Dense models : Qwen3‑32B, Qwen3‑14B, Qwen3‑8B, Qwen3‑4B, Qwen3‑1.7B, Qwen3‑0.6B

Flagship Performance

The flagship MoE model Qwen3‑235B‑A22B achieves competitive results on code, mathematics, and general‑ability benchmarks, matching or surpassing leading models such as DeepSeek‑R1, o1, o3‑mini, Grok‑3, and Gemini‑2.5‑Pro.

Reasoning Modes

Qwen3 supports two distinct inference modes that let users trade off latency against depth of reasoning:

Thinking mode – The model performs step‑by‑step reasoning, producing carefully considered answers suitable for complex problems.

Non‑thinking mode – The model returns a fast, near‑instant response for simpler queries where speed is prioritized. // 多种思考模式 This dual‑mode design enables explicit control of the model’s “thinking budget” per request.

Multilingual Support

All eight models understand 119 languages and dialects, providing broad coverage for international applications.

Agent Capabilities

Qwen3 is optimized for tool‑calling agents. In the BFCL (Benchmark for Foundation‑model‑based Agents) evaluation, Qwen3 attains a score of 70.8, surpassing Gemini‑2.5‑Pro, OpenAI‑o1, and other top models, thereby lowering the barrier for agent‑driven tool use.

The models natively support the MCP (Model‑Centric Protocol) and integrate with the Qwen‑Agent framework, which includes ready‑made tool‑calling templates and parsers.