What is this?
A family of open-source multimodal chat models that understand images, video and text, positioned as an open alternative to GPT-4o. InternVL2.5-78B was the first open-source model to top 70% on the MMMU benchmark.
A family of open-source multimodal chat models that understand images, video and text, positioned as an open alternative to GPT-4o. InternVL2.5-78B was the first open-source model to top 70% on the MMMU benchmark.Is this repo the models themselves, or code around them?This repo is the hub: training, evaluation and deployment code plus a model zoo. The actual weights, from 1B up to 241B parameters, are downloaded via the HuggingFace or ModelScope links in the zoo.
This repo is the hub: training, evaluation and deployment code plus a model zoo. The actual weights, from 1B up to 241B parameters, are downloaded via the HuggingFace or ModelScope links in the zoo.How do I get it running for a first chat?Install the Python dependencies from requirements.txt per the Installation Guide, grab a checkpoint from HuggingFace, and use the transformers quick start — or launch the bundled Streamlit demo to chat locally.
Install the Python dependencies from requirements.txt per the Installation Guide, grab a checkpoint from HuggingFace, and use the transformers quick start — or launch the bundled Streamlit demo to chat locally.Any gotchas before I download a checkpoint?Checkpoints ship in two layouts: the legacy GitHub format and the official transformers-standard HF format. If a load fails, convert between them with the custom2hf and hf2custom scripts.
Checkpoints ship in two layouts: the legacy GitHub format and the official transformers-standard HF format. If a load fails, convert between them with the custom2hf and hf2custom scripts.Does everything run locally, or do my images go to a server?You download the weights and run them on your own hardware — the repo even ships a local Streamlit chat demo, so images never have to leave your machine.
You download the weights and run them on your own hardware — the repo even ships a local Streamlit chat demo, so images never have to leave your machine.How big are these models — will my GPU handle one?Sizes run from 1.1B to a 241B-A28B MoE, with mid-size MoE variants like 20B-A4B and 30B-A3B. Small 1B-8B models suit modest GPUs, and 4-bit AWQ quantized builds exist for the heavier ones.
Sizes run from 1.1B to a 241B-A28B MoE, with mid-size MoE variants like 20B-A4B and 30B-A3B. Small 1B-8B models suit modest GPUs, and 4-bit AWQ quantized builds exist for the heavier ones.What can it actually do with images and video?OCR and document or chart understanding (SOTA among open models on DocVQA and InfoVQA), video question answering, image classification, semantic segmentation, image-text retrieval, and agentic tasks in the 3.5 generation.
OCR and document or chart understanding (SOTA among open models on DocVQA and InfoVQA), video question answering, image classification, semantic segmentation, image-text retrieval, and agentic tasks in the 3.5 generation.What is the newest model in the family?InternVL3.5, released August 2025. Its largest model, InternVL3.5-241B-A28B, is state of the art among open-source MLLMs on general multimodal, reasoning, text and agentic tasks; a 20B-A4B version is built on GPT-OSS-20B.
InternVL3.5, released August 2025. Its largest model, InternVL3.5-241B-A28B, is state of the art among open-source MLLMs on general multimodal, reasoning, text and agentic tasks; a 20B-A4B version is built on GPT-OSS-20B.Besides the weights, what else has been open-sourced?Training code for CascadeRL (an offline plus online RL pipeline) with the MMPR-v1.2 and MMPR-Tiny preference datasets, the VisualPRM 8B process reward model with VisualPRM400K data, and ShareGPT-4o: 200K images, 10K videos and 10K audios with descriptions.
Training code for CascadeRL (an offline plus online RL pipeline) with the MMPR-v1.2 and MMPR-Tiny preference datasets, the VisualPRM 8B process reward model with VisualPRM400K data, and ShareGPT-4o: 200K images, 10K videos and 10K audios with descriptions.Can I fine-tune it on my own data?Yes — every generation ships finetune guides, including a LoRA tutorial for COCO captioning, and the MPO training pipeline with its data-construction scripts is public too.
Yes — every generation ships finetune guides, including a LoRA tutorial for COCO captioning, and the MPO training pipeline with its data-construction scripts is public too.What are similar projects?ZDTaichu5.0-9B, jeeves, MiniCPM-V, DeepSeek-OCR, halo, Transformers-Tutorials, UniVR, EvoVLM
ZDTaichu5.0-9B, jeeves, MiniCPM-V, DeepSeek-OCR, halo, Transformers-Tutorials, UniVR, EvoVLMWhich topics is it in?gpt-4o, semantic-segmentation, open-weights, vlm, vision-language-model, reasoning, image-classification, transformers, ocr, gpt, reinforcement-learning, quantization
gpt-4o, semantic-segmentation, open-weights, vlm, vision-language-model, reasoning, image-classification, transformers, ocr, gpt, reinforcement-learning, quantization