Why Qwen3 Is Redefining Open‑Source LLMs: Mixed‑Inference Power and Unmatched Performance

In recent days open‑source large models have been extremely active. Rumors about DeepSeek‑R2 surfaced, Qwen3 was briefly announced and then withdrawn, and finally the model was officially released, positioning Alibaba ahead of DeepSeek in the emerging LLM arms race.

Qwen3 Overview

Qwen3 is Alibaba’s newest generation in the Qwen series and the first domestic “mixed‑inference” model. Mixed‑inference integrates top‑tier inference models with non‑inference models into a single architecture, requiring highly refined and innovative design and training. Apart from Qwen3, only Claude 3.7 and Gemini 2.5 Flash achieve similar capabilities.

Performance Explosion

The flagship model Qwen3‑235B‑A22B outperforms leading models such as DeepSeek‑R1, o1, o3‑mini, Grok‑3 and Gemini‑2.5‑Pro across coding, mathematics, and general‑ability benchmarks.

Additionally, the smaller MoE model Qwen3‑30B‑A3B surpasses QwQ‑32B with ten times the activation parameters, and even the micro‑model Qwen3‑4B rivals the performance of Qwen2.5‑72B‑Instruct.

Key Capabilities

Seamless switching within a single model between thinking mode (for complex logical reasoning, mathematics, and coding) and non‑thinking mode (for efficient general conversation).

Multilingual support covering 119 languages and dialects, with strong instruction‑following and translation abilities.

Enhanced agent capabilities: supports MCP protocol and custom tool integration, improving writing abilities in both modes.

Extreme cost control: the full‑version can be deployed on just four H20 GPUs.

Extended context length up to 128 K tokens.

Quick Start

Currently Qwen3 is available on ollama and openrouter, allowing developers to quickly integrate and experiment with the model.

Finally, what do you think about the upcoming DeepSeek‑R2—can it surpass Qwen3? Share your thoughts in the comments.