Our last posts drew a line from two models. A fair reader asked: is that a pattern, or two points? So we counted every fine-grained MoE model in the atlas — and the answer is sharper than we expected.
In the last two posts we noticed that MLA — DeepSeek’s low-rank attention compression — looked concentrated in Chinese labs. But we drew that from a handful of examples, and a sharp reader pushed back on exactly the right thing: two data points aren’t a pattern.
Fair. So we did the thing a benchmark leaderboard can’t: we counted. Every fine-grained Mixture-of-Experts model in the ModelAtlas catalogue, and for each one, a single question — does it use MLA, or not?
The naive answer is a mirage
Group the fine-grained MoE models by the country of the lab that made them, and MLA adoption looks like this: China 26%, United States 20% — nearly the same. You’d conclude MLA is spreading evenly across both frontiers. That conclusion is completely wrong.
You’d be wrong, and the reason is the whole point of ModelDNA. That 20% is 13 models, and every one is an NVIDIA NVFP4 re-quantization of a Chinese base — repackagings of DeepSeek, GLM, and Kimi. Not a single US-designed architecture among them. They have MLA because they are Chinese models in a 4-bit wrapper. Inheriting an architecture is not adopting it.
Strip the re-quants and count only original architectures — models a lab actually designed — and the picture snaps into focus.
The real picture: US original MLA adoption is zero
Zero. As of this count (July 2026), not one US lab has shipped an original architecture that uses MLA. The 20% from a moment ago was all 13 NVIDIA re-quants of someone else’s model.
A word on our own small samples, since this whole post is a lesson in not over-reading them: 26 US, 8 French, and 3 Korean original MoE models is thin. A 0-of-26 US count doesn’t prove no American will ever ship MLA — it rules out MLA being common in US original designs (statistically, the upper bound sits around 12%). The findings the numbers firmly support are the two that rest on real sample size: the China signal (41 of 135) and the China-vs-US gap.
It’s a five-lab club — and even China isn’t unanimous
Look at which labs ship original MLA models and the concentration is stark. The entire list is five Chinese labs plus one European outlier:
- DeepSeek (17 models) · Zhipu / GLM (10) · Moonshot / Kimi (8) · inclusionAI (5, Ant Group) · meituan-longcat (4, Meituan) — all Chinese. The lone non-Chinese original is Mistral’s 2026
Mistral-Small-4-119B(France) — a single model, the exception that proves the rule.
Here’s the nuance that makes this credible rather than a flag-wave: MLA isn’t even universal inside China. It’s 30% of Chinese MoE models, not 100%. Alibaba’s Qwen — one of the largest labs in the world — ships fine-grained MoE and stays on grouped-query attention. So the honest claim isn’t “Chinese labs use MLA.” It’s “five Chinese frontier labs use MLA, and almost no one else does.”
The contrast that tells the story
Now put it next to the other half of DeepSeek’s playbook. The DeepSeek-style MoE recipe — many fine-grained experts plus an always-on shared expert — is everywhere: every country, every size class, hundreds of models. It has become table stakes.
Two ideas came out of the same lab. One (the MoE recipe) went universal. The other (MLA) stayed a five-lab specialty. That divergence — same origin, opposite diffusion — is the actual finding, and it’s invisible unless you read architectures across the whole field.
Why only this kind of analysis surfaces it
Two things had to be true to get the real number, and both are ModelDNA’s core:
- Read architecture across the entire corpus — not benchmarks, not model cards, the actual attention and MoE structure of every model.
- Be strict about provenance — counting NVIDIA’s re-quants as “US MLA” is exactly the error that turns a real 0% into a fake 20%. Telling adoption from inheritance is the whole game.
No benchmark table encodes either. It lives in the weights, and in the lineage.
Paste any public HuggingFace model ID into the ModelDNA scanner — it tells you the architecture and whether a model is original or inherited, free and in seconds.
Counts are over fine-grained MoE models (≥32 experts) in the ModelAtlas catalogue with a current analysis; MLA is detected from the low-rank KV-compression signal in the config; “original” means no declared base model (re-quants and finetunes excluded). The catalogue is large but not a census of every model in existence.