whatmodel.sh / LLaVA-OneVision Qwen2 72B (ov-sft)
LLaVA-OneVision Qwen2 72B (ov-sft)
↗ share
+ contribute
LLaVA-OneVision Qwen2 72B (ov-sft) is an open-weights language model from LLaVA (lmms-lab), released August 2024. Its published size is 73.2B parameters. Licence: apache-2.0.
Catalogue identity notes
Comparison-only. Arena entry 'llava-onevision-qwen2-72b-ov' (vision). LMArena's leaderboard table (arena-leaderboard space @6f2eeeb) labels 'llava-onevision-qwen2-72b-ov' as 'LLaVA-OneVision-qwen2-72B-ov-sft' (LLaVA, Apache 2.0), linking https://huggingface.co/lmms-lab/llava-onevision-qwen2-72b-ov. Maker repo lmms-lab/llava-onevision-qwen2-72b-ov-sft @1ac2142: licence tag apache-2.0. 73,180,522,016 parameters from the Hugging Face safetensors metadata (dense, so active = total). Released 2024-08: month the maker's Hugging Face repo was created (createdAt 2024-08-06). LMArena labels the key 'LLaVA-OneVision-qwen2-72B-ov-sft'; its link lmms-lab/llava-onevision-qwen2-72b-ov redirects to lmms-lab/llava-onevision-qwen2-72b-ov-sft.
$ how it compares
The LLMs ranked either side of LLaVA-OneVision Qwen2 72B (ov-sft), out of 828 with a score.
| # | model | score | where |
|---|---|---|---|
| 537 |  Command R Plus (04-2024) | 45±1 | local |
| 538 |  Mistral NeMo Instruct (2407) | 44±1 | local |
| 539 | LLaVA-OneVision Qwen2 72B (ov-sft) | 44±3 | local |
| 540 |  Phi-3-mini-4k-instruct (June 2024 update) | 44±1 | local |
| 541 |  Yi 1.5 9B Chat | 44±1 | local |
$ what it's good at
vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #138 of 147 · 96% of its score · 3 tests
Trails Qwen3.5 0.8B, Molmo 7B-D (0924), North Micro Vision Instruct (2.4B) here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Arena vision (style controlled) | 978.3 ✎ | ▁▁▃▂▂▂▄▅▅▆██ beats 6/145 | 51% |
| CapArena-Auto detailed image captioning (Score_Avg, GPT-4o judge) | -12.3300 ✎ | ▃▃▃·▃█·▅▃▃▃▃ beats 4/12 | 35% |
| MSEarth-Bench Earth-science figure MCQ (Overall accuracy) | 51.1% ✎ | █·██····██·█ beats 2/5 | 10% |
$ where to run it
Other ways to run it
- Run it as published · needs ~146.36 GB, NVIDIA only · from github.com/vllm-project
vllm serve llava-hf/llava-onevision-qwen2-72b-ov-hf
Official weights and documentation: huggingface.co/lmms-lab ↗
$ how the score was worked out · LMArena leaderboard dataset 51%, CapArena: Benchmarking and An… 35%, MSEarth: A Multimodal Benchma… 10%
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 LMArena leaderboard dataset | 1 | 51% |
 CapArena: Benchmarking and Analyzing Detailed Image Captioning in the LLM Era (arXiv:2503.12329v1) | 1 | 35% |
| MSEarth: A Multimodal Benchmark for Earth Science Phenomenon Discovery with MLLMs (arXiv:2505.20740v3) | 1 | 10% |
| prior · pull toward the middle | 4% | |
| weighted average | = 44.3 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json