I saw a benchmark table on X showing a new model called Qwen3.6-35B-A3B outperforming Claude Sonnet-4.5 on almost every vision and reasoning benchmark.
My first reaction was disbelief. My second reaction was to immediately check if I could run it on my laptop.
What Is Qwen3.6-35B-A3B and Why Does It Matter
The name tells you everything once you decode it. 35B means 35 billion total parameters. A3B means only 3 billion are activated per token. This is Mixture of Experts architecture, the model has the knowledge capacity of a 35B model but runs at the compute cost of roughly a 3B model. Released by Alibaba’s Qwen team in April 2026, it’s a multimodal model handling text, images, and video, specifically designed for agentic coding and reasoning tasks. The benchmark showed it beating Claude Sonnet-4.5 on many things.
Step 1 : Downloading the Model
I already had Ollama installed and Open WebUI running at localhost:8080. I opened PowerShell and ran: ollama run qwen3.6:35b-a3b-q4_K_M. The download started immediately, 23GB total, about 10–15 minutes. When it finished, the CLI prompt appeared and I was talking to the model.
Step 2 : First Benchmark
Before touching any settings, I ran the model with the –verbose flag and sent a short prompt. Results: 16.32 tokens/sec eval rate, 45.06 tokens/sec prompt eval, 121ms load time, 2048 default context. 16 tokens per second for a 23GB model on a laptop. I was not expecting that.
Step 3 : Quality Test
I asked a real question I use professionally: “What is the difference between AI governance and AI compliance? Answer as a senior consultant.” The thinking trace showed the model building a delivery framework before writing a single word, deconstructing both concepts, adopting the consultant persona deliberately, self-correcting mid-draft. Then the actual response came: aviation analogy, “checkbox AI” framing, RACI mention, governance-as-floor-not-ceiling positioning, a comparison table, and a three-step recommendation. Good quality response.
Step 4 : Configuring Optimal Settings
The default context of 2,048 tokens is dangerously low for a model with thinking mode, the thinking trace itself consumes tokens before the answer starts. I needed to configure via Modelfile because Ollama’s CLI doesn’t support –option flags (Mistake 2) and PowerShell uses backtick not for line continuation (Mistake 1). The Modelfile approach bakes settings permanently into a named model: context window 32,768, temperature 1.0, top_p 0.95, top_k 20, presence_penalty 1.5, repeat_penalty 1.0, max_tokens 8,192. After configuration: 17.32 tokens/sec, faster than default, because baking settings lets Ollama allocate resources more efficiently at load time. I had to test this multiple times to confirm.
Step 5 : One-Click Launcher
My previous startup required opening Ollama, PowerShell, finding a saved command, copying it, pasting it, then navigating to localhost:8080. I built a .bat file to do it all with one double-click. Three more mistakes along the way: the standard Desktop path doesn’t exist when OneDrive is active and Windows is in Spanish (Mistake 3); RAG environment variables caused Open WebUI to crash (Mistake 4); the & background operator is bash syntax and doesn’t work in .bat files (Mistake 5). The final working file: start Ollama, wait 10 seconds, launch Open WebUI in its own cmd window, wait 30 seconds, open Chrome to localhost:8080.
What I Actually Learned
MoE architecture is not a compromise, it’s a genuine breakthrough for running capable models on consumer hardware. The thinking mode changes the character of the output significantly. And yes, I made five documented mistakes getting here. Every single one was fixable in under five minutes A 23GB model, launched with a double-click, producing governance consulting output that competes with a frontier cloud model. In April 2026, on consumer hardware. That’s where we are.