Technical Overview
InternVL3_5‑14B is the 14‑billion‑parameter flagship of the InternVL3.5 series, an open‑source family of multimodal large language models (MLLMs) that can understand and generate text conditioned on images. Built on top of the OpenGVLab/InternVL3_5‑14B‑MPO base checkpoint, this model is fine‑tuned with a novel Cascade Reinforcement Learning (Cascade RL) pipeline that first stabilises the model with offline RL and then refines its alignment with online RL. The result is a model that excels at complex visual‑reasoning tasks while remaining highly efficient.
- Key capabilities: image‑to‑text generation, visual question answering, multimodal dialogue, GUI interaction, and embodied agency.
- Architecture highlights: a decoupled vision encoder‑language model design (DvD), a Visual Resolution Router (ViR) that dynamically selects token resolution, and a two‑stage Cascade RL training loop that boosts reasoning performance by up to +16 %.
- Intended use cases: multimodal assistants, visual tutoring, document analysis, e‑commerce product search, and any application requiring deep visual‑language reasoning.
Benchmark Performance
InternVL3_5‑14B is evaluated on a comprehensive suite of multimodal benchmarks that cover general understanding, reasoning, text‑only, and agentic tasks (e.g., MMBench v1.1, MMStar, HallusionBench, MathVista, MMMU, MVBench, VideoMME, MME‑RealWorld, etc.). The README reports an average gain of +16 % over its predecessor and a 4.05× inference speedup, positioning it among the top open‑source MLLMs and narrowing the gap with commercial giants such as GPT‑5. These benchmarks are crucial because they test not only image captioning but also logical deduction, mathematical problem solving, and interactive agentic behaviours—areas where traditional vision‑language models often fall short.
- Compared to earlier InternVL3 models, the 14‑B variant delivers higher accuracy on MMMU and MathVista while maintaining faster throughput.
- Against competing open‑source models (e.g., LLaVA‑13B, MiniGPT‑4), InternVL3_5‑14B consistently outperforms on reasoning‑heavy metrics (MathVision, LogicVista) and matches or exceeds text‑only LLM baselines on MMLU‑Pro.
Hardware Requirements
Running InternVL3_5‑14B at full precision (FP16) requires roughly 28 GB of VRAM for the vision encoder + language model combined. For practical deployment, a GPU with ≥ 40 GB VRAM (e.g., NVIDIA A100 40/80 GB, RTX 4090 24 GB with tensor‑parallelism) is recommended to accommodate the dynamic token resolution of ViR and to achieve the reported 4.05× speedup. CPU load is modest; a modern 8‑core processor suffices for preprocessing and token handling. Storage wise, the checkpoint (including safetensors) occupies about 30 GB on disk, plus additional space for the MMPR‑v1.2 and MMPR‑Tiny datasets (≈ 15 GB). Inference latency is typically < 200 ms per 512‑pixel‑wide image on a single A100, thanks to the DvD and ViR optimisations.
Use Cases
InternVL3_5‑14B shines in scenarios where visual context and sophisticated language reasoning intersect.
- Customer support chatbots that can view screenshots, product images, or UI mock‑ups and answer user queries with step‑by‑step guidance.
- Educational tools that solve math problems from handwritten notes or explain scientific diagrams in multiple languages.
- E‑commerce search that matches user‑uploaded photos to catalog items, generating rich product descriptions on the fly.
- Document processing for OCR‑rich PDFs, extracting tables, charts, and providing natural‑language summaries.
- Embodied agents that can interpret GUI screenshots and issue simulated mouse/keyboard actions, enabling autonomous software testing.
Training Details
InternVL3_5‑14B was fine‑tuned from the OpenGVLab/InternVL3_5‑14B‑MPO checkpoint using a two‑stage Cascade RL pipeline. The offline RL phase stabilises the model on a large corpus of image‑text pairs, while the online RL phase refines alignment with human‑in‑the‑loop feedback. Training data includes the OpenGVLab/MMPR‑v1.2 (≈ 2 M image‑text pairs) and the smaller MMPR‑Tiny (≈ 200 k pairs) datasets, covering multilingual captions, visual QA, and instruction‑following examples. Compute was performed on a cluster of 8 × NVIDIA A100 80 GB GPUs for roughly 150 k steps, consuming an estimated 1.2 M GPU‑hours. The model retains the ability to be further fine‑tuned on domain‑specific data via the standard transformers Trainer API, with LoRA or full‑parameter updates.
Licensing Information
The model card lists the Apache‑2.0 license, which is a permissive open‑source licence. This licence grants you the right to use, modify, distribute, and even commercialise the model, provided you retain the original copyright notice and include a copy of the licence. No royalties are required, but you must provide attribution to OpenGVLab and include a notice of any modifications you make. The “unknown” tag in the meta‑data is a placeholder; the official README clarifies the Apache‑2.0 terms, so you can safely incorporate the model into commercial products, research pipelines, or SaaS offerings.