Why Qwen3 Is the New Powerhouse in Open‑Source AI Models

Qwen3‑235B‑A22B is one of the strongest open‑source models, achieving competitive results on coding, mathematics, and general benchmarks compared with top models such as DeepSeek‑R1, o1, o3‑mini, Grok‑3, and Gemini‑2.5‑Pro.

The small MoE model Qwen3‑30B‑A3B outperforms QwQ‑32B while using only one‑tenth of the activation parameters; even the tiny Qwen3‑4B rivals the performance of Qwen2.5‑72B‑Instruct.

Unique support for seamless switching between a "thinking" mode (for complex reasoning, math, and coding) and a "non‑thinking" mode (for efficient general dialogue) via the commands

/think

and

/no_think

in user prompts or system messages.

Significantly enhanced reasoning ability, surpassing previous QwQ (thinking mode) and Qwen2.5‑Instruct (non‑thinking mode) in mathematics, code generation, and commonsense logic.

Superior human‑preference alignment, delivering more natural, engaging, and immersive conversations in creative writing, role‑play, multi‑turn dialogue, and instruction following.

Outstanding agent capabilities: precise integration of external tools in both thinking and non‑thinking modes, achieving leading open‑source performance on complex agent‑based tasks and native support for the MCP protocol.

Supports over 100 languages and dialects (officially 119), providing strong multilingual instruction following and translation, including low‑resource languages such as Javanese and Haitian Creole.

Training data comprises 360 trillion tokens—double that of Qwen2.5—sourced from the web, extensive PDF extraction, and large amounts of synthesized code.

Deployment requirements are dramatically reduced; the flagship model can run locally on just four H20 GPUs, about one‑third the resources needed for the R1 model.

Model List and Execution

Running these models requires Ollama 0.6.6 or later.

<code>ollama --version
ollama version is 0.6.6</code>

0.6B Parameter Model

<code>ollama run qwen3:0.6b</code>

8B Parameter Model

<code>ollama run qwen3:8b</code>

14B Parameter Model (14.8B, Q4_K_M quantized, ~9.3GB)

<code>ollama run qwen3:14b</code>

32B Parameter Model

<code>ollama run qwen3:32b</code>

30B Mixture‑of‑Experts Model (3B activation parameters)

<code>ollama run qwen3:30b-a3b</code>

235B Mixture‑of‑Experts Model (22B activation parameters)

<code>ollama run qwen3:235b-a22b</code>

OpenRouter Full‑Version Free Access

<code>curl https://openrouter.ai/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -d '{
    "model": "qwen/qwen3-235b-a22b:free",
    "messages": [
      {"role": "user", "content": "What is the meaning of life?"}
    ]
  }'</code>

Inference Mode

Provides a soft‑switch mechanism that allows users to dynamically control model behavior with /think and /no_think . When enable_thinking=True , the model follows the most recent instruction in multi‑turn dialogue.

Non‑Inference Mode

Illustrates the model operating without the thinking mode enabled.

Conclusion

The author tested the full‑blood Qwen3‑235B‑A22B model in PIG AI’s function‑calling and JSON scenarios, noting that the results are subjective. In complex chart‑processing tasks such as Chat2BI, Qwen3’s capabilities were slightly weaker than DeepSeek V3‑0324, with average instruction‑following performance. The test used OpenRouter’s API, representing an initial experience; further evaluation is expected after Alibaba’s Bailei optimizations.