InternLM provides a suite of open-source large language models (LLMs) including InternLM, InternLM2, InternLM2.5, and InternLM3. These models are designed for general-purpose usage, advanced reasoning, and efficient deployment, offering state-of-the-art performance on various benchmarks. The project targets researchers and developers looking for high-performance LLMs with flexible deployment options.

How It Works

InternLM models are transformer-based architectures trained on massive, high-quality datasets. The latest InternLM3-8B-Instruct model, for instance, was trained on 4 trillion tokens, achieving competitive performance with significantly reduced training costs. It supports a "deep thinking" mode using long chain-of-thought for complex reasoning and a standard response mode for conversational interactions.

Quick Start & Requirements

• Installation: Primarily through Hugging Face Transformers (pip install transformers), LMDeploy (pip install lmdeploy), SGLang (pip install sglang), Ollama (ollama pull internlm/internlm3-8b-instruct), or vLLM (pip install vllm --pre --extra-index-url https://wheels.vllm.ai/nightly).
• Prerequisites: Python >= 3.8, PyTorch >= 1.12.0 (2.0.0+ recommended), Transformers >= 4.38. GPU acceleration is highly recommended for inference.
• Resources: InternLM3-8B-Instruct can be loaded in 4-bit precision, requiring approximately 8GB of GPU memory.
• Documentation: HuggingFace, LMDeploy, SGLang, Ollama, vLLM.

Highlighted Details

• InternLM3-8B-Instruct claims superior performance over Llama3.1-8B and Qwen2.5-7B on reasoning and knowledge-intensive tasks.
• Models support both standard inference and a "Thinking Mode" for complex reasoning tasks.
• A wide range of model sizes are available, from 1.8B to 20B parameters, including base, chat, and reward models.
• Long-context capabilities are supported, with InternLM2.5-Chat-1M handling up to 1 million tokens.

Maintenance & Community

• Active development with frequent releases of new models and updates.
• Community channels include Discord and WeChat.
• Links to issue reporting and technical reports are provided.

Licensing & Compatibility

• Code and model weights are licensed under Apache-2.0.
• The license permits commercial application.

Limitations & Caveats

• Despite safety efforts, models may produce unexpected outputs, including biases or harmful content. Users are advised against propagating such content.
• Evaluation results may vary due to OpenCompass version updates.