Alibaba has just released Qwen3.6 under an open-source license. 35 billion total parameters. 3 billion active parameters per inference. Benchmark results: > SWE-bench Verified: 73.4. Compared to Google’s Gemma4-26B-A4B at 17.4. > Terminal-Bench 2.0: 51.5. Google’s dense Gemma4-31B scores 42.9. > GPQA Diamond: 86.0. MMLU-Pro: 85.2. In computer vision, Alibaba claims parity with Claude Sonnet 4.5 on most multimodal benchmarks, with superiority in spatial reasoning. Do you grasp the magnitude of this? We’re talking about an OPEN SOURCE, LOCAL model approaching the performance of much larger and more expensive models. A model with only 3B active parameters runs on commodity hardware. A single consumer-grade GPU with proper quantization can serve the entire model. No cluster needed. No H100 required. No CapEx budget necessary. And it delivers coding agent performance comparable to dense models ten times larger in active parameters. This completely changes the math. For those building on closed APIs today, the question has shifted. It’s no longer whether open models will catch up to closed ones in capability. It’s how long until self-hosting a 3B-active-parameter model becomes cheaper than the API cost of an equivalent closed provider. The answer just got much shorter.