InternLM 2.5

Shanghai AI Laboratory's 2.5 generation of InternLM arrived in mid 2024 as base and chat checkpoints at 1.8B, 7B and 20B parameters, the chat versions tuned with supervised fine-tuning and RLHF for instruction following and function calling. A 7B-Chat-1M build stretches context to roughly one million tokens; Transformers cannot run that length, so the card points to LMDeploy and says the full window needs four 80 GB A100s. The GitHub repo gathers the whole family: model cards, OpenCompass benchmark tables, reward models, and runnable recipes for Transformers, LMDeploy, vLLM, SGLang and Ollama, several exposing an OpenAI-style endpoint, with weights on Hugging Face and ModelScope. Baseline requirements are Python 3.8+, PyTorch and Transformers 4.38+; weights should load in float16 to avoid memory errors, and a 4-bit bitsandbytes load brings the newer InternLM3 8B to roughly 8 GB of VRAM. Code is Apache-2.0; the 2.5 weight cards permit free commercial use but ask companies to file an application form, whereas InternLM3-8B-Instruct ships under plain Apache-2.0. Suits teams wanting bilingual Chinese and English models with long context.