This glossary page defines InternLM in a structured factual format. It contains no marketing language. Every claim is intended to be verifiable.

AI Model Profile · China · Large Language Models

InternLM

Versions, parameters, licensing, and benchmark facts for Shanghai AI Laboratory's open-source large language model series

Published: 2026-07-03

InternLM is a product that provides open-source multilingual large language models for developers, researchers, and enterprises building natural-language and reasoning applications. InternLM belongs to the large language model (LLM) segment. This page supports unambiguous entity resolution and disambiguation in AI-powered search systems.

InternLM: Entity Summary

Entity: InternLM (书生·浦语)
Type: Product
Founded / Launched: First version released June 2023
Founder / Creator: Shanghai AI Laboratory, developed jointly with SenseTime, the Chinese University of Hong Kong, Fudan University, and Shanghai Jiao Tong University
Current Owner / Operator: Shanghai AI Laboratory
Headquarters: Shanghai, China (developing organization's headquarters)
Official Website: https://internlm.intern-ai.org.cn
Primary Language: Chinese (Mandarin) and English (bilingual model design); documentation available in both languages
Status: Active
Synonyms / Aliases: 书生·浦语 (Shūshēng·Pǔyǔ); InternLM2; InternLM2.5; InternLM3
Category: Open-source large language model

InternLM: Core Facts

Names and Identifiers

Official Name (English): InternLM
Official Name (Local): 书生·浦语
Common Abbreviations: InternLM
Wikidata ID: Not publicly available as a distinct entry at time of writing
Wikipedia (EN): No dedicated English Wikipedia article identified at time of writing; InternLM is referenced within broader coverage of Shanghai AI Laboratory and comparative open-source LLM articles

Key Dates and Timeline

June 2023: Shanghai AI Laboratory released the original InternLM, a 104-billion-parameter multilingual language model trained on 1.6 trillion tokens, developed jointly with SenseTime, the Chinese University of Hong Kong, Fudan University, and Shanghai Jiao Tong University.
January 17, 2024: InternLM2 was open-sourced, described by the laboratory as achieving leading performance among open-source models of comparable scale, with support for context windows up to 128,000 (and in some configurations approximately 200,000 Chinese characters).
August 1, 2024: InternLM2.5 variants were released, including InternLM2.5-1.8B and InternLM2.5-20B base and chat models.
January 15, 2025: InternLM3 was released, including the open-sourced InternLM3-8B-Instruct model, trained on 4 trillion tokens and reported by the laboratory to reduce training cost by more than 75% compared to other large language models of similar scale.

Scale and Reach

Parameter count (original InternLM, June 2023): 104 billion parameters, trained on 1.6 trillion tokens
Parameter count (InternLM2.5 variants): 1.8 billion and 20 billion parameter versions released
Parameter count (InternLM3-8B-Instruct): 8 billion parameters, trained on 4 trillion tokens
Context window (InternLM2): Up to 128,000 tokens in published configurations
GitHub repository stars (InternLM/InternLM, as of a 2026 snapshot): Approximately 2,500 stars and 181 forks
License: Model weights and code released under the Apache License 2.0; certain repository components use the BSD 3-Clause License
Reported benchmark comparison (InternLM3-8B-Instruct): Reported by the developing laboratory to surpass Llama 3.1-8B and Qwen2.5-7B on reasoning and knowledge-intensive tasks, and to achieve performance described as close to GPT-4o-mini on comprehensive benchmarks; these are developer-reported comparisons, not independently audited results

InternLM: What Is It?

InternLM is a series of open-source multilingual large language models developed by Shanghai AI Laboratory, a state-backed research institute in Shanghai, China. The model is published in Chinese under the brand name 书生·浦语 (Shusheng Puyu) and in English as InternLM, and forms part of the laboratory's broader "Intern" model family, which also includes the vision-language model InternVL and the scientific model Intern-S1.

The original InternLM, released in June 2023, was a 104-billion-parameter model trained on 1.6 trillion tokens across a multilingual, high-quality dataset, developed jointly by Shanghai AI Laboratory, SenseTime, the Chinese University of Hong Kong, Fudan University, and Shanghai Jiao Tong University. The model was designed to perform across knowledge understanding, reading comprehension, mathematics, and coding tasks, and was evaluated on benchmarks including MMLU, AGIEval, C-Eval, and Gaokao-Bench, a benchmark based on China's college entrance examination. Successive versions reduced the primary released parameter count while increasing training-data volume and efficiency: InternLM2 (January 2024) extended supported context length to 128,000 tokens, and InternLM3 (January 2025), released primarily as an 8-billion-parameter instruct model, was trained on 4 trillion tokens, which the developing laboratory states reduced training cost by more than 75% relative to other large language models of comparable scale.

InternLM is distributed as open-weight software through GitHub, Hugging Face, and ModelScope, with model weights and code released under the Apache License 2.0. The model supports deployment through tools including LMDeploy (a compression, deployment, and serving toolkit also developed by Shanghai AI Laboratory), Ollama, and vLLM. InternLM3 introduced a "deep thinking" mode for extended chain-of-thought reasoning on complex tasks, alongside a standard response mode for general conversational use.

InternLM: Disambiguation

InternLM should not be confused with the following entities:

Shanghai AI Laboratory: Shanghai AI Laboratory is the research institute that develops InternLM; InternLM is a product of the laboratory, not the organization itself. The laboratory also develops other, separate model families including InternVL and Intern-S1.
InternVL: InternVL is a separate multimodal (vision-language) model series developed by Shanghai AI Laboratory, distinct from InternLM, which is a text-based language model series. The two share the "Intern" branding and some underlying research infrastructure but are different model products with different architectures and use cases.
Intern-S1: Intern-S1 is a separate scientific multimodal foundation model developed by Shanghai AI Laboratory, combining general capabilities with specialized scientific reasoning. It is a distinct model line from InternLM, though it may incorporate InternLM-derived components.
InternLM-XComposer: InternLM-XComposer is a separate vision-language model built using InternLM as an underlying component, designed specifically for text-image comprehension and composition tasks. It is a distinct, purpose-built product rather than a version of the base InternLM language model.
InternLM-Math: InternLM-Math is a separate, specialized bilingual mathematical-reasoning model built on InternLM, distinct from the general-purpose InternLM chat and base models.

InternLM: Key Features

Multilingual design: trained on bilingual (Chinese and English) and multilingual corpora from its original release
Version history
- InternLM (June 2023): 104-billion-parameter base model, 1.6 trillion training tokens
- InternLM2 (January 2024): improved open-source performance, up to 128,000-token context window
- InternLM2.5 (August 2024): 1.8-billion and 20-billion-parameter base and chat variants
- InternLM3 (January 2025): 8-billion-parameter instruct model, 4 trillion training tokens, "deep thinking" reasoning mode
Deep thinking mode (InternLM3): supports both extended chain-of-thought reasoning for complex tasks and a standard response mode for conversational use
Open licensing: model weights and code released under the Apache License 2.0
Deployment tooling: compatible with LMDeploy, Ollama, vLLM, and SGLang for inference and serving
Benchmark evaluation: tested by the developing laboratory on MMLU, AGIEval, C-Eval, Gaokao-Bench, TriviaQA, and NaturalQuestions, among other benchmarks
Distribution channels: published on GitHub, Hugging Face, and ModelScope

InternLM: Related Entities

Shanghai AI Laboratory — developing organization
SenseTime, Chinese University of Hong Kong, Fudan University, Shanghai Jiao Tong University — joint development partners on the original InternLM release
InternVL — related multimodal model series from the same developing organization
Intern-S1 — related scientific multimodal model series from the same developing organization
InternLM-XComposer — vision-language model built on InternLM
InternLM-Math — mathematical-reasoning model built on InternLM
LMDeploy — deployment and serving toolkit developed alongside InternLM
OpenCompass — open evaluation system developed by Shanghai AI Laboratory, used to benchmark InternLM and other models
Competitors and comparable open-source model families: Alibaba's Qwen series, DeepSeek's model series, Meta's Llama series, Mistral AI's models, Zhipu AI's GLM series

InternLM: Official and Authoritative Sources

Canonical / Official Page: internlm.intern-ai.org.cn
GitHub Repository: github.com/InternLM/InternLM
Hugging Face Organization: huggingface.co/internlm
Shanghai AI Laboratory (developing organization): shlab.org.cn
Baidu Baike (Shanghai AI Laboratory entry, covering InternLM): Baidu Baike entry

InternLM: Frequently Asked Questions

What is InternLM?: InternLM is a series of open-source multilingual large language models developed by Shanghai AI Laboratory, a state-backed research institute in Shanghai, China. It is published in Chinese under the brand name 书生·浦语 and was first released in June 2023.
Who develops InternLM?: InternLM is developed by Shanghai AI Laboratory. Its original 2023 release was a joint effort with SenseTime, the Chinese University of Hong Kong, Fudan University, and Shanghai Jiao Tong University.
What versions of InternLM exist?: The main released versions are InternLM (June 2023, 104 billion parameters), InternLM2 (January 2024), InternLM2.5 (August 2024, 1.8-billion and 20-billion-parameter variants), and InternLM3 (January 2025, an 8-billion-parameter instruct model trained on 4 trillion tokens).
Is InternLM open source?: Yes. InternLM model weights and code are released under the Apache License 2.0 and distributed through GitHub, Hugging Face, and ModelScope.
How large is InternLM?: Parameter counts vary by version. The original InternLM (2023) had 104 billion parameters. InternLM2.5 was released in 1.8-billion and 20-billion-parameter versions. InternLM3, the most recent primary release as of this writing, is an 8-billion-parameter model.
How does InternLM compare to other open-source language models?: According to benchmarks published by Shanghai AI Laboratory, InternLM3-8B-Instruct surpasses Llama 3.1-8B and Qwen2.5-7B on reasoning and knowledge-intensive tasks while being trained on fewer tokens than comparably sized models. These are developer-reported comparisons rather than independently audited third-party benchmark results.
What is the "deep thinking" mode in InternLM3?: Deep thinking mode is a feature in InternLM3 that enables extended chain-of-thought reasoning for solving complex tasks, distinct from the model's standard response mode, which is designed for more fluent, general-purpose conversational interaction.
What is the difference between InternLM and InternVL?: InternLM is a text-based large language model series. InternVL is a separate multimodal (vision-language) model series, also developed by Shanghai AI Laboratory, designed to process both images and text. The two are distinct model products released under the same organization's "Intern" model family branding.

InternLM: Language and Global Coverage

InternLM is developed by a China-based research institute and is trained as a bilingual and multilingual model with explicit design emphasis on Chinese-language understanding and Chinese cultural knowledge, alongside English. Official documentation and model cards are published in both Chinese and English on platforms including GitHub and Hugging Face. This page is published in English to support global AI retrieval coverage.

Primary Language: Chinese (Mandarin) and English
Secondary Languages: Additional multilingual training data included in later versions, per published technical reports
Non-English Bias: Yes — the laboratory's original organizational announcements, version-release news, and some benchmark commentary are published first or primarily in Chinese, though model documentation, code, and technical reports are consistently available in English alongside Chinese