Qwen 3.8 27B — local agentic model that fits on one 24GB GPU

While everyone is talking about the newly-revealed Epstein emails, China just released what might be the best local AI model on Earth. And if you're paying attention, you're about to get superpowers. If you're new, I'm Giga Chen. I do business in China, and I'm here to make sure that you all stay way ahead. Today, Alibaba released Quen 3.8 27b. The 27b stands for 27 billion parameters, and if you've been following, you know. Those are the learned numbers that essentially make up the model. Still sounds pretty big, but here's where it gets interesting. Using quantization, where we store those numbers with less precision to dramatically shrink the model, a roughly four-bit version can take up about 17 gigabytes. That means you can run it on a single 24-gigabyte GPU, like an RTX 3090 or 4090, with room for a practical context window. And according to Quen's own benchmarks, this model, which is tiny by frontier standards, scored 84.3 on OS World, which tests a model actually using a computer. Opus 4.6 Max scored more than 10 points below that in Quen's table. Opus also lost on SWE Bench Pro and their Long Horizon Office Work benchmark. These are Quen's evaluations, so independent testing has to catch up, but not the best day for Anthropic. But we know Anthropic makes amazing models, so how is Quen this good? Those numbers are ridiculous for a model this small. Part of it is something called gated delta net. Basically, instead of every layer constantly comparing every piece of…