InternVL 3.5
Shanghai AI Lab's open vision-language family (书生·万象), nine sizes from 1B to 241B-A28B under Apache 2.0. The default open VLM for OCR, document parsing and GUI agent work.
Overview
Nine checkpoints on a single day, 26 August 2025, running from 1B to 241B-A28B. That release pattern is what makes the InternVL family useful, more than any individual benchmark row. The same architecture and the same training recipe scale from a 1B model on an edge device up to a mixture-of-experts flagship, so a prototype that works at 4B does not need its prompting and post-processing rebuilt when it graduates to 38B.
The architecture is the familiar ViT-MLP-LLM stack. The 8B pairs InternViT-300M with Qwen3-8B for 8.5B total parameters. The larger mixture-of-experts variants at 20B-A4B, 30B-A3B and 241B-A28B are initialized from GPT-OSS, which tends to surprise people who assume a Shanghai lab builds only on domestic foundations. Shanghai AI Lab reports a 16.0% overall reasoning improvement and a 4.05x inference speedup against InternVL3, helped by a Visual Resolution Router that halves visual token count in the Flash variants with little accuracy cost.
Two things decide whether this fits a China deployment. Weights sit on ModelScope as well as Hugging Face, and vllm-ascend has supported InternVL3.5 since v0.20.2 with MindSpeed-MLLM handling training, so this is one of a very small number of frontier-class open vision-language models with a real path onto Huawei Ascend silicon. That is not a footnote if you are selling into a state-owned enterprise. The counterweight is momentum: the main repository's last major version is a year old, and while the team shipped InternVL-U, a 4B unified understanding, generation and editing model, on 6 March 2026, a proper InternVL4 has not landed.
Key Features
- ✓ Nine sizes shipped together on 26 August 2025: 1B, 2B, 4B, 8B, 14B and 38B dense models plus mixture-of-experts variants at 20B-A4B, 30B-A3B and 241B-A28B.
- ✓ ViT-MLP-LLM architecture. The 8B pairs InternViT-300M with Qwen3-8B for 8.5B total; the large MoE variants are initialized from GPT-OSS rather than a domestic base.
- ✓ A Visual Resolution Router in the Flash variants cuts visual token count by up to 50% with close to unchanged accuracy.
- ✓ Shanghai AI Lab reports a 16.0% overall reasoning gain and a 4.05x inference speedup over InternVL3.
- ✓ Supported in vllm-ascend from v0.20.2, with MindSpeed-MLLM covering distributed training on Huawei Ascend NPUs.
- ✓ The MIT-licensed repository ships full training, fine-tuning and evaluation code, not just an inference reference.
Where it holds
- • Apache 2.0 weights across the whole family including the 241B flagship, with no community licence, no user cap and no excluded regions.
- • The size ladder is real. Same architecture and same recipe from 1B on an edge box up to 241B-A28B, so scaling up does not mean rewriting your prompting and post-processing.
- • One of very few frontier-class open vision-language models with a documented, tested Ascend deployment path, which decides the project when your customer cannot buy NVIDIA.
- • Strong on OCR and layout parsing for Chinese documents, invoices and forms, which is where most imported VLMs quietly fall over.
Where it breaks
- • No major version in twelve months. InternVL3.5 dates from August 2025 and the team shipped InternVL-U, a 4B unified model, on 6 March 2026, but an InternVL4 has not appeared and a year is a long gap in this field.
- • Weights inherit terms from their base models, so verify per size rather than assuming one blanket licence across the family.
- • Serving 241B-A28B needs a multi-node setup. For most teams the practical sweet spot is the 8B dense or the 30B-A3B MoE.
- • English documentation is solid for inference and thin for the training pipeline, and a fair number of configuration questions only get real answers in Chinese issue threads.
My Take
Nine checkpoints released on the same day in August 2025, from 1B up to a 241B-A28B mixture of experts, every one of them Apache 2.0. The detail most write-ups miss is that the large MoE variants are initialized from GPT-OSS rather than a domestic base, which is not what people assume about a Shanghai lab. For OCR and document work on Chinese layouts it is the first thing I would reach for, and vllm-ascend has supported it since v0.20.2 if your deployment cannot use NVIDIA. The honest worry is the calendar: nothing major since August 2025.
Quick Info
- Pricing:
- open-source
- Openness:
- Open weights
- Licence:
- Apache 2.0 (repository code is MIT)
- Starting at:
- Free. All nine InternVL3.5 checkpoints are downloadable from Hugging Face and ModelScope under Apache 2.0, commercial use included, with no registration and no user-count clause. The repository code, including training and evaluation pipelines, is MIT. There is no first-party paid API, so serving cost is whatever your own hardware or cloud bill comes to.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where InternVL 3.5 holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to InternVL 3.5
MiniCPM
open-sourceOpenBMB's on-device line (面壁小钢炮): MiniCPM5-1B at 1.08B parameters with a 131K context, plus MiniCPM-V 4.6 multimodal at 1.3B. Apache 2.0 on weights and code.
Google Gemma 4
open-sourceGoogle's open-weight family in five sizes, 2B effective up to 31B dense, text and image in with audio on the smaller ones, 256K context and Apache 2.0 weights since April 2026.
InternLM
open-sourceShanghai AI Lab's open-weight family (书生·浦语), now the Intern-S scientific multimodal line: 35B up to 1T parameters, Apache 2.0, trained on domestic computing infrastructure.
Meta Muse Glimmer
open-sourceMeta's 30B dense multimodal agent model. Apache 2.0, 131K context, 4-bit checkpoints under 20GB and none of the Llama licence conditions attached.