InternLM
By Shanghai AI Laboratory
InternLM is a family of open-source large language models developed by Shanghai AI Laboratory in collaboration with SenseTime, the Chinese University of Hong Kong, and Fudan University, spanning general-purpose chat, reasoning, and…
Definition
InternLM is a family of open-source large language models developed by Shanghai AI Laboratory in collaboration with SenseTime, the Chinese University of Hong Kong, and Fudan University, spanning general-purpose chat, reasoning, and long-context variants.
Overview
InternLM began as a research project from Shanghai AI Laboratory (also known as Shanghai AI Lab), one of China's leading state-backed AI research institutions, and has since grown into a widely used open-weight model family. The original InternLM and InternLM-7B/20B models were released in 2023, followed by InternLM2 in early 2024, which improved reasoning, long-context handling (up to 200K tokens in some variants), and tool use. InternLM2.5 and InternLM3 continued this trajectory, adding stronger math and coding performance and more efficient training recipes, with InternLM3-8B in particular emphasizing strong performance relative to its parameter count. The project is tightly coupled with an ecosystem of companion tools: XTuner for fine-tuning, LMDeploy for efficient inference and quantization, OpenCompass for evaluation, and Lagent/AgentLego for building tool-using agents. This full-stack approach — models plus training, deployment, and evaluation tooling — mirrors what Hugging Face and Meta's Llama ecosystem offer, and is a deliberate strategy to make InternLM a practical foundation for both research and production use in China and internationally. InternLM models are released under permissive licenses (Apache 2.0 for code and many model weights) and are available on Hugging Face and ModelScope. They compete directly with other Chinese open-weight families such as Qwen, Baichuan, and ChatGLM, as well as with Meta's Llama series, and have been benchmarked competitively on Chinese- and English-language reasoning, math, and coding tasks. The lab has also published InternLM-XComposer for multimodal document understanding and generation, extending the family beyond pure text.
Key Concepts
- Open-weight release under Apache 2.0 for most model checkpoints
- Long-context variants supporting up to 200K tokens
- Companion tooling: XTuner for fine-tuning, LMDeploy for inference, OpenCompass for evaluation
- Strong math and coding benchmark performance for its size class in later versions
- Built-in support for agent and tool-calling workflows via Lagent/AgentLego
- Multimodal extension through InternLM-XComposer for document and image understanding
- Available on Hugging Face and ModelScope with multiple parameter sizes
- Developed with academic partners CUHK and Fudan University