LLaVA became the fastest-rising open model family in post-ChatGPT AI research

A reproducible survey scanned 26,104 accepted CVPR, ICLR and NeurIPS papers from 2023–2025. Among six openly released model families reported in its VLM table, LLaVA rose from 0.1% to 2.7% of VLM abstracts—the clearest gain.

26,104
papers scanned
3
post-ChatGPT years
6
open families charted
+2.6 pp
LLaVA gain, 2023–25
Share of VLM abstracts referencing each model family · percent

Fractions are within the survey’s VLM-abstract subset, not all 26,104 papers. Families can co-occur, so lines do not sum to 100%.

What changed

01
LLaVA broke away from the pack. Its share climbed 27-fold, from 0.1% in 2023 to 2.7% in 2025. arxiv.org
02
MoE also accelerated. The architecture family moved from 0.6% to 1.3%, with the increase concentrated in 2025. arxiv.org
03
Earlier BLIP families receded. BLIP-2 fell from 0.6% to 0.2%, while BLIP slipped from 0.4% to 0.1%. arxiv.org
04
The broader open-weight trend is larger than VLM alone. A separate full-text study of 21 million scientific articles reports open-weight use reaching 44.0% of single-family papers in 2026. arxiv.org

Model-family fractions

Family202320242025ChangeSource
LLaVA0.1%1.2%2.7%+2.6 pparxiv.org
MoE0.6%0.6%1.3%+0.7 pparxiv.org
CLIP / OpenCLIP0.1%0.3%0.2%+0.1 pparxiv.org
GroundingDINO0.0%0.2%0.2%+0.2 pparxiv.org
BLIP-20.6%0.2%0.2%−0.4 pparxiv.org
BLIP0.4%0.1%0.1%−0.3 pparxiv.org

Method · One 26,104-paper survey covering accepted CVPR, ICLR and NeurIPS papers, 2023–2025, was used for the time series; its Table 1 measures the share of VLM abstracts referencing each family. A separate 21-million-article S2ORC study was used only for broader context. Accessed 24 August 2026.

Keenable · made with SELECT* · 2,027 pages in 2m 48s · Open the chatShare: XLinkedInReddit