AILANTA
← All research
Hugging Face · GitHub

Chinese open AI is forming its own ecosystem

We compared five Chinese model families with Llama, Mistral, and Gemma to test whether model releases are becoming a distinct layer of derivatives, infrastructure, and applications.

19.07.20262,684 Hugging Face655 GitHubProduct Hunt excluded
Main finding

This is no longer isolated model export, but it is not yet an evenly mature ecosystem

Chinese families produced 1,830 observed models and 486 GitHub integrations. Applications account for 31.7% of repositories versus 23.7% in the control cohort. Yet usage is concentrated around Qwen and DeepSeek, while the long Kimi and MiniMax tail consists mainly of derivatives with low median traction.

01

The ecosystem extends beyond models

GitHub cohorts by repository creation month, using the same minimum three-star threshold.

0275380106JanFebMarAprMayJun
China cohortControl cohort
GitHub projects486 vs 169

5 Chinese / 3 control families

Application layer31.7%

control: 23.7%

Median stars/day1 vs 0.8

current age-normalized snapshot

Derivatives and formats97.4%

non-base releases

02

Model supply is expanding in waves

Current top-by-downloads snapshot restricted to H1 2026 creation dates.

0122244366488JanFebMarAprMayJun
China cohortControl cohort

Absolute counts are not a census of the Hugging Face Hub: each family search is capped at 1,000 current results. The chart describes the observed cohort, not the entire market.

03

Five Chinese families are developing differently

Current downloads and stars indicate maturity, not historical growth.

FamilyModelsGitHubMedian downloads
Qwen38618751.4K
DeepSeek241145850
Kimi5269515
MiniMax6204236
GLM57170

Qwen combines a broad derivative layer with high median usage. DeepSeek has fewer model artifacts but the highest median GitHub stars/day. Kimi and MiniMax are proliferating across formats and fine-tunes, but much of that supply still receives little usage.

04

The control cohort shows a different profile

FamilyModelsGitHubMedian downloads
Llama53794.3K
Mistral15126518
Gemma650645K

The control cohort is more infrastructure- and local-distribution-heavy. This is not a model-quality ranking: it compares ecosystem shape around model families, not capability.

05

What this means for the market

01

The Chinese stack already has an application layer

Nearly a third of linked GitHub projects are applications, agents, and workflows rather than only runtimes and converters.

02

Qwen and DeepSeek remain the maturity centers

Cohort breadth does not imply even adoption: the two leaders have markedly stronger traction than the long tail.

03

Independence remains partial

Distinct derivatives and applications are forming, while local formats and deployment practices remain shared with global open-source infrastructure.