StudioThinking Machines releases open-weight multimodal Inkling 2 Small (594B MoE)
Original title: thinkingmachines/Inkling-2-Small is released!
Thinking Machines released the open-weight multimodal model Inkling 2 Small: a 48-layer MoE transformer with 594B total and 17B active parameters, text and image inputs, and a 1M-token context window. The NVFP4 checkpoint needs at least 500GB aggregated VRAM (e.g., 2x NVIDIA B300), runs via SGLang, vLLM, TokenSpeed, or Unsloth, and the card includes broad benchmark comparisons plus safety evaluations concluding no material uplift beyond the existing open-weight ecosystem.
The model card details the architecture, NVFP4/MXFP8 hardware requirements, and extensive benchmark comparisons, letting readers gauge deployment cost and performance.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.