internvl

A Next-Generation Training Engine Built for Ultra-Large MoE Models

[ICCV 2025] Are VLMs Ready for Autonomous Driving? An Empirical Study from the Reliability, Data, and Metric Perspectives

Higher performance OpenAI LLM service than vLLM serve: A pure C++ high-performance OpenAI LLM service implemented with GPRS+TensorRT-LLM+Tokenizers.cpp, supporting chat and function call, AI agents, distributed multi-GPU inference, multimodal capabilities, and a Gradio chat interface.

[ICLR 2026] OptMerge: data-free merging of multimodal LLM experts (VQA, Geometry, Chart, OCR, Grounding) and modalities, plus SWUDI / ASWUDI closed-form spectral merging for CLIP-ViT, Flan-T5, Llama-3.2 and MLLMs.

PriorTR (ECCV 2026): training-free, prior-corrected visual token reduction for accelerating multimodal LLMs — image & video.

[ICML 2026 Oral] DOUBT: Decoupled Object-level Understanding and Bridging via vMF-based Trustworthiness for Hallucination Detection in MLLMs

[ICML 2026] RLSF-V: Mitigating Hallucinations in MLLMs via Fuzzy Semantic Self-Feedback

Streamlit App Combining Vision, Language, and Audio AI Models

VLM (InternVL3) inference on Jetson Orin Nano (8GB) using TensorRT-LLM.

Automate Microsoft Account OAuth registration and token imports with this Chrome extension for the Microsoft Account Manager API.

Fine-tuned sub-1B vision-language models that read CV/resume page images like a human and output ATS-ready JSON. No PDF text layer, runs on a laptop CPU.

Unified evaluation harness for Vision-Language Models — VQAv2, TextVQA, COCO, MMBench, MMBench-CN

天择多模态决策大模型框架 · TianZe-MJev: A Multimodal Jev-style Decision Model。七个官方基座、免额外训练、统一 API 与网页决策台。

[ACM AI 2026] Safety-audit framework for vision-language models on brain MRI, measuring calibration, confident errors, hallucination, abstention, and slice-level accuracy.

帧防 FrameGuard · 视觉安全智能分析 — 用多模态大模型逐帧理解监控画面，对照自定义规则判定违规并输出分时段报告。更多炸裂应用尽在知识星球“AI创收私研社”

Description: Domain-shift failure analysis of LLaVA-1.5-7B and InternVL2-8B across photos, charts, medical, and screenshots. 945 probes, 6 failure categories. Chart OCR: LLaVA 0% vs InternVL2 71%. Domain-dependent failure confirmed via chi-square (p<0.0001).

ACL 2025, SDP workshop Shared Task Submission

上海人工智能实验室（SHLab）开源项目精选导航 · 含许可证风险标注与失效链接勘误

Unofficial PyTorch reproduction for InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models.

Same image, different answer: open vision-language models (Qwen3-VL, MedGemma, InternVL3.5) change their mammogram readings when only the prompt changes. Pre-specified study, 103,500 model calls, fully reproducible.