I'm using the qwen3 4B model in consumer hardware, which beats Gemini in English language tasks.