Index-1.9B

Why we included this project

Small models earn their keep when GPU budgets are tight, and Bilibili's Index-1.9B line targets exactly that situation. The family spans a base model, a chat variant aligned with SFT and DPO, a role-playing edition that adds RAG for few-shot character customization, and a 32K long-context version that can handle documents over 35,000 words. Trained mainly on Chinese and English, the chat model stands out among peers of its size for a livelier conversational style and strong translation between East Asian languages. Because the weights are available through llama.cpp and Ollama, you can run it on consumer hardware for chatbots, document reading, or character-driven demos without provisioning a large serving stack. That makes it a practical pick for small teams who want a multilingual model they can actually afford to run.