InternVL¶

Development line: project:internvl · thread internvl-model-releases
Events: 2 dated, 2024-12-11 → 2025-08-27 · Researched: 2026-09-04 · confidence: high

What it is¶

InternVL is an open-weight vision-language model family for teams that need to analyse images, documents, multiple images, video and GUI states in a self-hosted workflow.

• Image, document, chart and OCR question answering
• Multi-image and video conversation
• Visual grounding, multilingual vision-language work, plus GUI and embodied tasks in 3.5

The current 3.5 line ranges from 1.1B to 240.7B total parameters; the official guide places models through 30B on one A100, 38B on two, and 241B on eight. Choose a complete no-suffix 3.5 checkpoint for new local work; use Flash when visual-token cost is the bottleneck.

Development line¶

• 2024-12-11 — InternVL2.5-78B model and collection released. On 2024-12-11, the project released InternVL2.5-78B with a model collection, repository, project site, and demo. The linked sources do not establish benchmark results or feature details.
• 2025-08-27 — InternVL3.5 paper and model collection released. On 2025-08-27, the project published the InternVL3.5 paper, official collection, repository, and chat. The linked sources do not establish the paper's detailed claims, supported variants, or deployment guidance.

What changed¶

On 2024-12-05 (recorded 2024-12-11), InternVL2.5 became a 1B–78B family. It kept the earlier core architecture but revised training, test-time scaling, and data-quality work. The 78B checkpoint paired a 6B vision encoder with Qwen2.5-72B.

On 2025-08-26 (recorded 2025-08-27), InternVL3.5 replaced that core line with 1B–241B checkpoints in both project-native and standard Hugging Face formats. It introduced CascadeRL, the Visual Resolution Router, decoupled vision-language deployment, Qwen3/GPT-OSS language backbones, and GUI/embodied capabilities.

On 2025-08-30, the project opened CascadeRL training code and data for its GPT-OSS 20B-A4B variant. After 2025-08-27, an official InternVL3.5-Flash checkpoint became available. Its card documents ViCO plus visual-resolution routing to reduce visual tokens by 50%; the reviewed sources did not establish its exact first release date.

How to use this¶

On 2024-12-11, evaluate the official InternVL2.5-78B model page and collection when selecting an InternVL2.5 release. By 2025-08-27, review the linked InternVL3.5 paper and official collection before picking a newer InternVL line instead of assuming the earlier line remains the default.

1. Choose a complete no-suffix InternVL3.5 checkpoint for normal inference; choose the Flash variant when visual-token cost matters. — https://huggingface.co/OpenGVLab/InternVL3_5-8B-Flash
2. Use the format deliberately: the project-native checkpoint loads through AutoModel with remote code, while the HF format follows standard Transformers conventions. Install transformers >=4.52.1; the 20B-A4B variant needs >=4.55.0. — https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct
3. For image chat, apply the supplied 448px dynamic tiling preprocessing, convert the tiles to bf16, and send the prompt with an placeholder through model.chat. — https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct
4. For video, sample a bounded number of frames, prefix each with FrameN: , and pass the matching num_patches_list to the chat call. — https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct
5. For a service, deploy through LMDeploy or vLLM with an OpenAI-compatible endpoint; use vLLM for the GPT-OSS 20B-A4B variant. — https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct

Best practices¶

• For deployment, prefer the no-suffix checkpoint: it is the version that completed the full pipeline, unlike Pretrained, Instruct and MPO stage checkpoints. — https://huggingface.co/OpenGVLab/InternVL3_5-8B-Flash
• Plan model size from the documented GPU envelope before downloading: through 30B on one A100, 38B on two, and 241B on eight. — https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct
• Use the documented Transformers version floor; the 2.5-era dependency floor is not sufficient for 3.5. — https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct
• For thinking mode, set do_sample=true and temperature=0.6; the project recommends this to limit repetition. — https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct
• Increase the context window for multi-image prompts and number images in the prompt so the model can distinguish them. — https://huggingface.co/OpenGVLab/InternVL3_5-8B-Instruct

Superseded by this¶

• 2024-12-05 / 2024-12-11: InternVL2.5 is no longer the default for a new core InternVL deployment; use a complete no-suffix InternVL3.5 checkpoint unless legacy compatibility is the deciding constraint.
• The InternVL2.5 transformers>=4.37.2 guidance is obsolete for InternVL3.5: use transformers>=4.52.1, or >=4.55.0 for 20B-A4B.
• 2025-08-26: the 3.5 card's guidance that Flash was still forthcoming is obsolete; an official InternVL3_5-8B-Flash checkpoint is now available.

Still unknown¶

• The exact first publication date of InternVL3.5-Flash was not established; current availability is confirmed, but a model-page update date is not necessarily a launch date.
• The reported reasoning gains and token reduction are project-reported benchmark results, not an independent evaluation for a particular workload.
• The public chat frontend was reachable but did not expose an inspectable response to the fetcher, so live hosted-demo behaviour was not verified.

Sources¶

Agent brief¶

• Subject: project:internvl, thread internvl-model-releases, 2 dated events 2024-12-11 → 2025-08-27.
• Practical note: As of 2024-12-11, practitioners should evaluate the official InternVL2.5-78B model page and collection when selecting an InternVL2.5 release. As of 2025-08-27, they should review the linked InternVL3.5 paper and official collection before choosing a newer InternVL line, rather than assuming the earlier model line remains the appropriate default.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.