InternLM2.5
Shanghai AI Lab's updated InternLM generation with stronger tool-use and reasoning
InternLM2.5 is Shanghai AI Laboratory’s follow-up to InternLM2, released starting July 3, 2024 in 7B and 1.8B sizes, with a 20B version added on August 1, 2024. The 7B chat model scores 72.8 on MMLU and beats Llama 3 8B Instruct and Gemma 2 9B on several reasoning and math benchmarks, according to the developer’s own OpenCompass evaluations. A dedicated 7B-Chat-1M checkpoint extends the context window to one million tokens and can locate specific facts inside million-token documents in needle-in-a-haystack tests, while the standard 20B model tops out at 200K tokens. The series also adds stronger function calling and agent-style tool use compared to the original InternLM2, and all weights are released under the Apache 2.0 license through Hugging Face and the InternLM GitHub repository.