Inkling 2 Small

Inkling 2 Small is a general-purpose multimodal model from Thinking Machines Lab that accepts text and image inputs and generates text outputs. It is a 48-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) feed-forward backbone – each token is routed to 8 of 384 experts – with a hybrid of local and global attention layers. The model is natively multimodal: images are encoded via a vision transformer and projected into a shared hidden space, processed jointly by the decoder.

From the model card: 594B total / 17B active parameters, context window of up to 1M tokens; pretrained, midtrained and post-trained in NVFP4 (NVFP4 is the native precision; an MXFP8 checkpoint variant exists). The NVFP4 checkpoint requires ≥500 GB aggregate VRAM (e.g. 2x NVIDIA B300 or 4x H200), the MXFP8 checkpoint ≥800 GB (e.g. 4x B300 or 8x H200). Released with open weights under Apache-2.0 to support research, fine-tuning and integration into third-party products. Intended for developers building agentic and tool-use systems, coding assistants, chatbots and retrieval-augmented generation systems, as well as general-purpose conversational use. Local inference via SGLang, vLLM, TokenSpeed or Unsloth; API access is available through NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare; fine-tuning is available via Tinker (Tinker Cookbook).

Training Data Broad mix of content types: text, images and video, drawn from publicly available sources, third-party acquired data, internally generated and synthetic/augmented data. Curation includes deduplication and filtering (Model Card §4). Documentation: https://thinkingmachines.ai/training-data-documentation/