PJLab

Open scientific pretraining corpus from OpenDataLab (Shanghai AI Lab ecosystem). 25M+ scientific documents / 600B+ tokens / 3.87 TB across 10 discipli…

Efficient scientific multimodal foundation model from InternLM, continued pretrained from Qwen3.5. Per the HuggingFace model card: 36B parameters (35B…

Landmark 1-trillion parameter MoE scientific model with 512 experts (8 active per token, 22B active parameters). Designed for autonomous scientific re…

Real-time interaction benchmark for Video LLMs.

A unified agentic framework for end-to-end, long-horizon autonomous scientific discovery — the rebrand and successor of NovelSeek ("NovelSeek has been…

A rethink of web-corpus construction that treats HTML-to-text extraction — usually a fixed preprocessing step — as a learned task. The pap…

Identifies a fundamental tokenization dilemma in Scientific LLMs processing biomolecular sequences: sequence-as-language (e.g., NatureLM, Intern-S1) a…

Visual-programmatic interface for code intelligence. Can "see" UI designs and "write" corresponding code.

Cross-platform computer use agent operating on Windows, macOS, Ubuntu, and Android. Sets SOTA on multiple GUI benchmarks including 94.4% on MMBench-GU…

Advances open-source multimodal models in versatility, reasoning, and efficiency. Introduces Cascade Reinforcement Learning (offline + online RL), Vis…

Scientific multimodal MoE foundation model with 241B total parameters (28B active). Trained on 5 trillion tokens including 2.5T+ from scientific domai…

Open-source framework of 1,000+ domain-diverse verifiable task environments for LLM reasoning research. Covers 8 domains including algorithms, cryptog…

Eliminates padding requirements in FP8 grouped GEMM on NVIDIA Hopper GPUs using a TMA descriptor pool approach. Achieves 1.7-20.4% speedup with up to…

Policy Discriminative Learning (POLAR) — a pre-training approach that frames reward modeling as distinguishing between different policies. Achieves si…

Multimodal benchmark probing scientific cognitive abilities across three levels: scientific signal perception, scientific attribute understanding, and…

A method for reinforcement learning on unlabeled data — training LLM reasoning at test time with no ground-truth rewards. The core insight: common tes…

Introduces native joint multimodal pre-training, variable visual position encoding, mixed preference optimization, and test-time scaling. InternVL3-78…

Proposes topic-based data partitioning as a replacement for the standard source-based approach to pretraining data mixing. Uses unsupervised clusterin…

The paper introducing MedXpertQA, a benchmark motivated by the observation that existing medical benchmarks have insufficient difficulty and clinical…

A challenging benchmark for expert-level medical reasoning and understanding: 4,460 questions spanning 17 specialties and 11 body systems, sourced fro…

Most advanced general-purpose LLM, focusing on system-level reasoning and self-reflection.

A comprehensive multilingual dataset focusing on "Silk Road" languages, containing over 300 billion tokens. It features major language subsets for Tha…

First open-source multimodal LLM to surpass 70% on MMMU. Investigates scaling of vision encoders, language models, datasets, and test-time Chain-of-Th…

Specialized reasoning model utilizing long-term thought chains and self-correction, similar to OpenAI's o1 paradigm.

Open-source multi-agent web search engine framework (similar to Perplexity).

Optimized for 1-million-token context windows and agentic tool usage.

First version to offer a full family of models from edge to frontier (1B, 2B, 4B, 8B, 26B, 40B, 76B, 108B). No standalone paper; technical details wer…

Benchmark for testing long-document understanding in multimodal models.

Introduced dynamic high-resolution processing (up to 4K via 1-40 tiles of 448x448), continuous learning for InternViT-6B, and high-quality bilingual t…

Groundbreaking multimodal dataset featuring 200K images and 10K videos with detailed, dense captions.

A 1.0T token high-quality English webtext dataset derived from Common Crawl, also known as WanJuan 2.0. Specifically designed for pre-training large l…

Significant upgrades in reasoning and math. Outperformed Llama-2-70B at release. Includes InternLM2-Math variant.

Foundational vision-language model that scaled ViT to 6B parameters (InternViT-6B) and aligned it with an LLM using web-scale image-text data. Achieve…

Introduces code-integrated mathematical reasoning where each solution interleaves natural language, code, and execution results. Created MathCodeInstr…

Specialized for advanced text-image composition, generating articles with interleaved images. Versions 1.0, 2.0, and 2.5.

Lightweight fine-tuning toolkit integrating core PEFT methods like LoRA and QLoRA.

Frameworks for building LLM-based agents that can use external tools and APIs.

Massive high-quality multimodal pre-training corpus containing over 2TB of English and Chinese text, image-text pairs, and video.

Large-scale video-text dataset for multimodal generation.

Debut 104B parameter foundational bilingual (Chinese/English) model.

Training framework designed for trillion-parameter models with minimal dependencies.

Industry standard for evaluating large models across thousands of dimensions.

Planning-oriented end-to-end autonomous driving framework unifying perception, prediction, and planning in one network. Won the CVPR 2023 Best Paper A…

High-performance toolkit for compressing, deploying, and serving LLMs at scale.

Large-scale video foundation models. InternVideo2 won multiple champions in the 2024 EgoVis challenges.

Vision foundation model utilizing deformable convolutions for high-performance detection and segmentation. CVPR 2023 Highlight.

World's first open-source decision intelligence platform, unveiled at WAIC 2021. DI-engine is the core generalized decision AI engine for PyTorch and…

OpenMMLab is the most influential open-source computer vision algorithm ecosystem, co-created by SenseTime, CUHK MMLab, and Shanghai AI Lab. Since Oct…