In this video, Sam Witteveen explores the capabilities of InternLM, specifically InternLM 2.5, a new language model optimized for math, reasoning, and function calling. Developed by the Shanghai AI Lab in collaboration with SenseTime, InternLM 2.5 is designed to support agentic tasks, making it a strong contender for building local agents and handling JSON well. The model has been topping the Hugging Face leaderboards and is noted for its robust performance in various benchmarks, including math and reasoning.
Sam highlights the model’s deployment framework, LMDeploy, which allows for efficient handling of long context windows by transferring data to disk for offline inference. Additionally, the video introduces Lagent, a lightweight framework built specifically for InternLM, designed to enhance tool use and function calling capabilities. The technical report for InternLM provides detailed insights into the data composition and training methodologies, emphasizing its focus on instruction following and tool selection.
The video includes a practical demonstration of InternLM’s performance using Hugging Face and Ollama implementations. Sam tests the model’s function calling abilities and showcases how it can be integrated into various applications. He also discusses the model’s unique chat format and its ability to handle function calls effectively. Overall, InternLM 2.5 is presented as a promising tool for developers looking to build sophisticated AI agents with strong reasoning and function calling capabilities.