Inkling-2-Small weights released (huggingface.co)

Inkling 2 Small has been released as a versatile multimodal model capable of processing both text and image inputs to generate text outputs. Designed for developers, this model supports various programming languages and is optimized for applications like coding assistants, chatbots, and tools that require natural language and visual comprehension. Its open weights enable easy integration, research, and fine-tuning by the community, significantly enhancing the development landscape for AI-driven applications. The technical architecture of Inkling 2 Small is robust, featuring a 48-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) backbone, allowing efficient routing of data. It natively supports inputs in both text (UTF-8) and image formats, maintaining up to a 1 million token context window. The model is available in two checkpoint formats—in NVFP4, which is recommended for most use cases, and the marginally improved MXFP8. This model's ability to function locally on GPU clusters with substantial VRAM opens doors for large-scale deployments while contributing to ongoing advancements in multimodal AI technologies.