Teaching giant models to pack light.

A Mixture-of-Experts model is a crowd of specialists. Qwen3-30B-A3B has 128 experts in each layer. For every token, a little router picks 8 of them and the rest sit the turn out.

That's why it only does about 3B params of work per token even though it's 30B big. Cheap to run, in theory. The catch: all 128 experts still have to live in memory, ready, in case the router calls on them.

The usual way to quantize is to give every weight the same precision, say 4 bits, and call it a day. But experts aren't equal. Some get called constantly and carry a lot. Others barely show up.

DuoRate leans into that. It allocates precision per expert, so the bits go where they actually change the output. I contributed to the research, and the headline result is the number in the panel's corner.

Next I built a post-training quantization pipeline for Qwen3-VL-30B, a vision-language model that reads images as well as text. "Post-training" means no retraining. You take the finished model and compress it.

The recipe was W2A16: 2-bit weights, 16-bit activations. Sixteen bits down to two is 8× smaller weights. That was the target, on Ascend 910B NPUs.

To spend bits wisely, you need to know which experts actually get used. So I dug into expert-routing behaviour across vision-language workloads: which experts light up for text, which for images, and how lopsided the traffic gets.

That map is what tells a compression scheme where precision is worth spending, and where it's wasted on an expert that barely gets called.

Not everything was bits. I also worked on TTD-DR, a deep-research agent that writes reports by searching, drafting, and revising. It kept searching for things it had basically already found, and it kept asking an LLM "is this relevant?" over and over.

I plugged in LightRAG, which builds a knowledge graph out of everything the agent has already read. Then I replaced the LLM relevance check with embedding-similarity gating: if a new query is close enough to something the graph already knows, skip the search.