Mentatcurated
Artificial Intelligence high · first-party

Memory that needs no math

Alibaba's new open model keeps a third of its weights in a plain lookup table of 20 million common word pairs, cheap enough to leave sitting in ordinary system memory instead of on the graphics chips.

Alibaba's Qwen team released the weights for Qwen3.8-Flash-Next, and of the roughly 180 billion parameters in the download, 51 billion are a lookup table: 20 million of the commonest two- and three-word sequences from its training data, each with a stored representation. The model consults that table once, near the start of processing each word, and moves on.

The table adds no arithmetic, so it does not have to sit where the arithmetic happens.

Looking something up costs no arithmetic. That is what makes the table odd to serve. It adds an enormous amount of stored knowledge without adding any work per word, and because a lookup is predictable, a server can fetch the entries it is about to need in advance from ordinary system memory rather than keeping them in the scarce, expensive memory bolted to the chips that do the arithmetic. Engineers running the model on a common eight-chip server measured what that buys: moving the table off the chips cut the weights each one had to hold from about 84 gigabytes to 60, and raised how much conversation the machine could keep in working memory by roughly three-quarters.

For most of the past decade, giving a language model more knowledge meant giving it more arithmetic to do. Mixture-of-experts broke that link once, by keeping a large pile of parameters around and firing only a small slice per word; this model keeps a 125-billion-parameter core and fires 6 billion of it at a time. The lookup table breaks the link a second way, with parameters that never fire at all. If the approach holds, how capable a model you can serve becomes partly a question of how much cheap RAM is in the box, not only what the accelerators can hold.

The idea is not Alibaba's. Feeding a model a large table of word-sequence embeddings has been circulating in research papers since early 2025; what is new is somebody shipping it at this size in a model anyone can download. Two claims to hold lightly: the company's headline figure, that this took about a ninth the training compute of its much larger sibling, is its own number from a comparison in which more than the architecture changed, and the memory-offload path currently runs only on Nvidia hardware. Qwen describes the release as a preview of the architecture its next generation will use, which is exactly how it framed the last model of this shape, and that one's ideas all shipped.

The lenses

Novelty 3
Impact · breadth 2
Impact · depth 3
Actionable 3
Delight 2
Substance 5
Hype 3

The facts

WeightsFree to download under Alibaba's community licence
What you needA multi-GPU server and a lot of host RAM; not a workstation model
ContextAbout 260,000 tokens at once, roughly two long novels
Released26 August 2026
Open huggingface.co →

How this connects

Tap a node to open it

How this connects