Model Releases
InternLM 3 8B Release: Deep Thinking & Apache 2.0
Shanghai AI Lab unveils InternLM 3, an 8B parameter model surpassing Llama 3.1 on reasoning tasks with 128K context and Apache 2.0 licensing.
March 5, 2025
Model ReleaseInternLM 3
Introduction
Shanghai AI Lab has officially announced the release of InternLM 3 on March 5, 2025, marking a pivotal moment for the open-source AI ecosystem. This new iteration represents a significant leap forward, offering an 8B parameter bilingual model specifically engineered for high-performance reasoning and complex logical deduction.
Unlike previous iterations, InternLM 3 introduces a specialized deep thinking mode that enhances problem-solving capabilities without sacrificing inference speed. For developers seeking a robust, cost-effective alternative to proprietary closed models, this release provides a compelling option with full Apache 2.0 licensing, ensuring freedom for commercial and academic use alike.
- Release Date: 2025-03-05
- Parameters: 8B
- License: Apache 2.0
Key Features & Architecture
The underlying architecture of InternLM 3 is built upon a robust foundation featuring a massive 128K context window, allowing for the seamless processing of extensive documents, long-form content, and complex data streams. Training data spans over 4 trillion high-quality tokens, ensuring the model possesses a vast and diverse knowledge base across both English and Chinese languages.
A key innovation is the reported 75% cost savings during the training phase, achieved through optimized data curation and highly efficient inference engines designed for modern hardware. This efficiency allows for faster deployment cycles and reduced operational overhead for organizations adopting the model for production environments.
- 8B Parameters
- Bilingual Support (English/Chinese)
- 128K Context Window
- Apache 2.0 License
- Deep Thinking Mode
Performance & Benchmarks
Independent benchmarks reveal that InternLM 3 surpasses Llama 3.1 8B and Qwen2.5 7B on critical reasoning and knowledge tasks, setting a new standard for open-source performance. On the MMLU benchmark, it achieves a score of 82.5, outperforming the previous best open-source 8B models by a significant margin.
In HumanEval, the model demonstrates superior code generation capabilities with a 78% pass rate, making it a strong contender for developer tools. Furthermore, SWE-bench results indicate strong problem-solving abilities in software engineering contexts, validating its utility for technical workflows and automated coding agents.