GITFIND

open-weight adoption — tracked daily

THE OPEN MODEL INDEX

Which open-weight models are developers actually building with? Hugging Face weight downloads and GitHub velocity for DeepSeek, Llama, Qwen, Kimi, Mistral, and GLM — plus the runtime layer that runs them. Every number says what it is; nothing here is made up.

30-day pulls: 5.8Mmodels tracked: 8 + 3 runtimesleader: kimi k3 1.6M/30dfastest riser: qwen3 235b · 31d

§ 1 — weight downloads, daily

fig. 01

5.0M10.0Mapr 30 — 4.7M commitsmay 2 — 154k commitsmay 7 — 795k commitsmay 8 — 4.7M commitsmay 9 — 463k commitsmay 10 — 580k commitsmay 11 — 795k commitsmay 13 — 693k commitsmay 15 — 396k commitsmay 16 — 947k commitsmay 19 — 4.9M commitsmay 21 — 5.3M commitsmay 23 — 592k commitsmay 25 — 4.4M commitsmay 27 — 4.7M commitsmay 29 — 138k commitsmay 30 — 5.0M commitsmay 31 — 1.0M commitsjun 1 — 682k commitsjun 2 — 328k commitsjun 4 — 7.4M commitsjun 6 — 5.7M commitsjun 7 — 5.7M commitsjun 8 — 5.7M commitsjun 11 — 6.0M commitsjun 15 — 218k commitsjun 16 — 100.0k commitsjun 17 — 558k commitsjun 18 — 204k commitsjun 19 — 996k commitsjun 21 — 6.8M commitsjun 24 — 7.0M commitsjun 29 — 696k commitsjun 30 — 171k commitsjul 1 — 8.8M commitsjul 3 — 8.0M commitsjul 4 — 151k commitsjul 5 — 745k commitsjul 7 — 1.1M commitsjul 9 — 8.6M commitsjul 10 — 11.1M commitsjul 11 — 175k commitsjul 12 — 8.7M commitsjul 14 — 142k commitsjul 15 — 10.8M commitsjul 17 — 229k commitsjul 18 — 917k commitsjul 19 — 69.4k commitsjul 20 — 8.6M commitsjul 22 — 10.1M commitsjul 24 — 208k commitsjul 26 — 130k commitsjul 28 — 195k commitsjul 29 — 11.4M commitsjul 30 — 11.4M commitsjul 31 — 11.9M commitsaug 1 — 12.2M commitsaug 2 — 12.0M commitsaug 3 — 12.5M commitsaug 4 — 12.6M commitsaug 5 — 13.1M commitsaug 6 — 13.1M commitsaug 7 — 13.0M commitsaug 8 — 12.8M commitsaug 9 — 12.6M commitsaug 10 — 12.3M commitsaug 11 — 12.2M commitsaug 12 — 12.1M commitsaug 13 — 12.1M commitsaug 14 — 12.3M commitsaug 15 — 12.2M commitsaug 16 — 12.0M commitsaug 17 — 11.7M commitsaug 18 — 11.5M commitsaug 19 — 11.5M commitsaug 20 — 11.3M commitsaug 21 — 10.8M commitsaug 22 — 10.3M commitsaug 23 — 10.1M commitsaug 24 — 9.8M commitsaug 25 — 9.7M commitsaug 26 — 9.6M commitsaug 27 — 8.5M commitsaug 28 — 7.7M commitsaug 29 — 7.7M commitsaug 30 — 7.3M commitsaug 31 — 7.0M commitssep 1 — 6.8M commitssep 2 — 6.8M commitssep 3 — 6.5M commitssep 4 — 5.9M commitssep 5 — 5.9M commitssep 6 — 5.9M commitssep 7 — 5.7M commitssep 8 — 5.6M commitssep 9 — 5.6M commitssep 10 — 5.6M commitssep 11 — 5.6M commitssep 12 — 5.5M commitssep 13 — 5.6M commitssep 14 — 5.5M commitssep 15 — 5.6M commitssep 16 — 5.7M commitssep 17 — 5.8M commitssep 18 — 5.8M commitssep 19 — 5.8M commitssep 20 — 5.8M commitssep 21 — 5.7M commitssep 22 — 5.7M commitssep 23 — 5.8M commitssep 24 — 5.8M commitssep 25 — 5.9M commitssep 26 — 5.8M commitssep 27 — 5.9M commitssep 28 — 5.8M commits5.8Mapr 30may 27jun 18jul 12aug 1aug 15aug 29sep 12sep 28

hugging face trailing 30-day weight downloads per day, across all tracked models.

§ 2 — share of new downloads

fig. 02

28%22%16%15%9%
kimi k3 28.3%deepseek v3 22.4%llama 3.3 70b 15.6%deepseek r1 15.4%mistral small 3.1 8.7%kimi k2 4.6%qwen3 235b 3.1%glm 4.5 1.8%

share of the trailing 30-day window.

§ 3 — each model, same scale

fig. 03

kimi k3

1.6M/30d · ▼ 38%

deepseek v3

1.3M/30d · ▲ 27%

llama 3.3 70b

905k/30d · ▲ 215%

deepseek r1

895k/30d · ▼ 73%

mistral small 3.1

507k/30d · ▲ 269%

kimi k2

270k/30d · ▲ 69%

qwen3 235b

182k/30d · ▲ 363%

glm 4.5

106k/30d · ▲ 9%

trailing 30-day weight downloads per day, averaged over 3-day buckets · all charts share one scale.

§ 4 — momentum, ranked

fig. 04

  • 01qwen3 235bfastest

    182k/30d · ▲ 363% · 31d to double · 3.1%

  • 02mistral small 3.1

    507k/30d · ▲ 269% · 39d to double · 8.7%

  • 03llama 3.3 70b

    905k/30d · ▲ 215% · 46d to double · 15.6%

  • 04kimi k2

    270k/30d · ▲ 69% · 93d to double · 4.6%

  • 05deepseek v3

    1.3M/30d · ▲ 27% · 196d to double · 22.4%

  • 06glm 4.5

    106k/30d · ▲ 9% · 546d to double · 1.8%

  • 07kimi k3

    1.6M/30d · ▼ 38% · — · 28.3%

  • 08deepseek r1

    895k/30d · ▼ 73% · — · 15.4%

§ 5 — the runtime layer

fig. 05

fig. 05 — the runtime layer (what people run open models with)

toolstarsforkscontributors
ollama182k18.0k596
llama.cpp130k23.8k2.0k
vllm92.8k22.7k2.4k

runtimes don’t publish weight downloads — repo traction is the honest proxy here.

§ 6 — intelligence brief

fig. 06

$ tail -f /var/log/gitfind/models.log

09:41:02 8 open models + 3 runtimes tracked — hf trailing-30d downloads + github velocity, nothing fabricated

09:41:03 kimi k3 leads the 30-day window at 1.6M

09:41:04 llama 3.3 70b gained the most rate this month — +618k on the 30d window

09:41:05 qwen3 235b is the fastest riser — doubling every 31 days at current pace

09:41:06 ollama leads the runtime layer at 182k★

09:41:07 router share coming when there’s a stable public source — we don’t scrape

09:41:08 █

form 27-b — subscription

THE TUESDAY BRIEFING

The repos that moved this week, why they matter, and what to watch next. One email. No noise.