CALM2-7B-Chat

CALM2-7B-Chat (CyberAgentLM2-7B-Chat) is a 7-billion-parameter bilingual dialogue model from CyberAgent, released in early 2024 under Apache 2.0. It is a chat fine-tune of CyberAgentLM2, a decoder-only model pretrained from scratch on 1.3 trillion tokens of publicly available Japanese and English data.

The chat variant extends the context window to 32,768 tokens and uses a structured USER/ASSISTANT chat template. As one of the first fully Apache-2.0 Japanese LLMs, it became a popular base for the Japanese open-source community, spawning vision-language derivatives such as llava-calm2-siglip and wide Ollama/GGUF distribution for local use.