A flagship on a gaming card

Alibaba released a model that activates 3 billion parameters for each token it generates. A year ago, the company's best model activated 22 billion to do comparable work. The new one keeps 35 billion in memory but lights up only a sliver at a time, and on Alibaba's own scorecard it lands roughly even with that older flagship — ahead on a broad knowledge test, behind on harder reasoning and some coding tasks. The headline isn't 'small beats big.' It's 'sparse rivals dense': the model isn't tiny, it's selective.

The catch: every benchmark here is Alibaba's own, and the model already lost the harder reasoning and coding tests to a flagship a year its senior.

That selectivity is the whole point. Because so few parameters fire per token, the model runs at over a hundred words a second on a single consumer graphics card costing around $800 — where the model it rivals needed datacenter-class memory to load at all. Frontier-adjacent coding and agent work that used to mean renting a server now fits on a desk, under an Apache 2.0 licence that lets anyone download and run it.

This isn't a breakthrough so much as the cost curve doing what it has been doing since DeepSeek and Mistral popularised the same trick: each generation, the parameters you actually pay to run drop while quality holds. The proof of how fast that curve moves is the model's own obsolescence — Alibaba shipped a whole newer generation within two months. The durable signal isn't this one release; it's that 'good enough to self-host' keeps arriving sooner than the hardware to run it gets cheaper.