LLaVA

weights

open(safetensors on the Hub, ungated)

data

open(LLaVA-OneVision-2-Data is published ungated under Apache 2.0, with the OneVision-1.5 mid-training and instruct sets carried forward)

code

open(training scripts and configs in EvolvingLMMs-Lab/LLaVA-OneVision-1.5 (examples/llava_onevision2))

license

Apache-2.0(OSI

LLaVA-OneVision-2 is released under Apache 2.0 with its training data, vision encoders, training code and logs, so the model can be reproduced from what ships.

• https://huggingface.co/api/datasets/lmms-lab/LLaVA-One-Vision-1.5-Mid-Training-85M recorded 2026-09-27

  Resolves to mvp-lab/LLaVA-OneVision-1.5-Mid-Training-85M: public, ungated, apache-2.0

• https://huggingface.co/api/datasets/mvp-lab/LLaVA-OneVision-2-Data recorded 2026-09-27

  LLaVA-OneVision-2-Data: public, ungated, apache-2.0

• https://huggingface.co/api/models/lmms-lab-encoder/LLaVA-OneVision-2-8B-Instruct?expand[]=downloads&expand[]=cardData&expand[]=gated&expand[]=createdAt&expand[]=lastModified&expand[]=siblings recorded 2026-09-27

  Hub metadata for LLaVA-OneVision-2-8B-Instruct: gated false, safetensors shards, card license apache-2.0, created 2026-05-11

• https://raw.githubusercontent.com/EvolvingLMMs-Lab/LLaVA-OneVision-1.5/HEAD/examples/llava_onevision2/quick_start_4b/quick_start.sh recorded 2026-09-27

  The LLaVA-OneVision-2 quick-start training script

• https://raw.githubusercontent.com/EvolvingLMMs-Lab/LLaVA-OneVision-1.5/HEAD/LICENSE recorded 2026-09-27

  The repository LICENSE is the Apache License, Version 2.0

• https://raw.githubusercontent.com/EvolvingLMMs-Lab/LLaVA-OneVision-1.5/HEAD/README.md recorded 2026-09-27

  "a fully open 8B multimodal model ... with the entire pipeline (data, encoders, training, checkpoints, logs) released end-to-end"; released 2026-04-30

Hugging Face downloads over the trailing 30 days, summed across the lmms-lab and lmms-lab-encoder checkpoints and Hugging Face's llava-hf conversions. Nearly all of it is the older LLaVA-1.5 and 1.6 conversions rather than the current OneVision releases.

• https://huggingface.co/api/models?author=llava-hf&search=llava&sort=downloads&direction=-1&limit=100 recorded 2026-09-27

  29 llava-hf conversions with 3,261,856 downloads in the trailing 30 days; llava-1.5-7b-hf 1,646,114

• https://huggingface.co/api/models?author=lmms-lab-encoder&limit=100 recorded 2026-09-27

  LLaVA-OneVision-2-8B-Instruct with 8,393 downloads in the trailing 30 days; the other five repositories are vision encoders

• https://huggingface.co/api/models?author=lmms-lab&search=llava&sort=downloads&direction=-1&limit=100 recorded 2026-09-27

  47 lmms-lab LLaVA checkpoints with 77,776 downloads in the trailing 30 days; LLaVA-OneVision-1.5-8B-Instruct 15,955

LLaVA-OneVision-2 reads single images, image sets and long video, and its predecessor already edged past Qwen2.5-VL-7B on documents and charts. It reports no GUI operation or audio input, so it sits level with MiniCPM-V.

• https://huggingface.co/lmms-lab-encoder/LLaVA-OneVision-2-8B-Instruct/raw/main/README.md recorded 2026-09-27

  "A multimodal vision-language model that handles single images, multi-image, and video inputs, built on a Qwen3-8B language backbone"; codec video backend

• https://huggingface.co/lmms-lab/LLaVA-OneVision-1.5-8B-Instruct/raw/main/README.md recorded 2026-09-27

  Evaluation Results: MMMU val 55.44, ChartQA 86.48, DocVQA test 95.00; Qwen2.5-VL-7B 51.33, 84.08, 94.93