2024 leader
Llama
32.4%
12 of 37 model–paper mentions in the targeted research sample.
Qwen overtook Llama in the 2026 research sample
A targeted scan of 1,499 arXiv search results found 175 model–paper records and 146 normalized model mentions from 2023 through 24 August 2026. Llama led the 2024–25 samples; in partial 2026, Qwen reached 29.9% while Llama fell to 17.9%.
29.9%
Qwen share, 2026 YTD
33.3%
Llama share, 2025
146
normalized model mentions
1,499
arXiv results scanned
2024
2025
2026
0%25%50%75%100%
What changed
1
Llama established the early default. It accounted for 32.4% of sampled model mentions in 2024 and 33.3% in 2025, spanning alignment, quantization, medicine and language adaptation. arxiv.org
2
Qwen rose fastest. Its share moved from no captured mentions in 2024 to 23.1% in 2025 and 29.9% in partial 2026. arxiv.org
3
The field diversified rather than converged. Gemma held 10.8–13.4% across 2024–26, while Phi, DeepSeek, Falcon, BLOOM, OPT and smaller families filled substantial niches. arxiv.org
4
2026 is not comparable as a full year. The latest slice stops on 24 August and may overrepresent newly searched Qwen-focused work. arxiv.org
Leaders by period
2025 leader
Llama
33.3%
13 of 39 mentions, with Qwen closing to 23.1%.
2026 YTD leader
Qwen
29.9%
20 of 67 mentions through 24 August, ahead of Llama and Mistral at 17.9% each.
Underlying fractions
| Family | 2024 | 2025 | 2026 YTD |
|---|---|---|---|
| Qwen | 0.0% | 23.1% | 29.9% |
| Llama | 32.4% | 33.3% | 17.9% |
| Mistral | 18.9% | 5.1% | 17.9% |
| Gemma | 10.8% | 12.8% | 13.4% |
| Phi | 8.1% | 12.8% | 4.5% |
| DeepSeek | 0.0% | 7.7% | 3.0% |
| Other | 29.8% | 5.2% | 13.4% |
Method: targeted web search across 1,499 arXiv result pages, sampled from AI/ML queries for major open-weight model families; 88 extracted paper records expanded to 175 model–paper rows, normalized to families and deduplicated by paper title. Fractions divide each family’s distinct paper mentions by all model mentions in that year, so a paper using several models contributes once to each. This is a directional research-use sample, not a complete census of arXiv. 2023 had only three captured Llama papers and is omitted from the main chart; 2026 runs through 24 August.