Qwen3 235B-A22B — In-Depth LOCAL Testing! (The BIGGEST Qwen Yet!)

Timestamps: 00:00 - Intro 01:30 - Local Settings 03:27 - First HTML Test 06:11 - Python Game Test 08:47 - No Thinking Mode 10:18 - JS Game Test 13:11 - JS Pong Game Test 15:55 - Second HTML Test 18:50 - VC Pitch Test 25:50 - Impressions & Closing Thoughts In this video, we take an in-depth look at the newly released Qwen3-235B-A22B MoE model, the largest model in the Qwen3 lineup. Touted as matching the performance of top-tier models like DeepSeek R1, OpenAI o1/03-mini, Grok, and Gemini 2.5, this Mixture-of-Experts model delivers strong results—especially considering its size and accessibility in quantized form. We go over the recommended sampling parameters to enable both “thinking” and “non-thinking” modes. From there, we test the model across a range of tasks including HTML generation, Python game development, JavaScript-based interactive outputs, and a JS pong game. We also walk through a VC pitch prompt and share general impressions of how this model feels in terms of reasoning, creativity, and usability when run locally. Hugging Face Repo: https://huggingface.co/Qwen