Article

Leading US open-weight model arrives: Inkling 2 Small requires at least 500GB of GPU memory

Beating AI reports that Thinking Machines, founded by former OpenAI CTO Mira Murati, released the weights for Inkling 2 Small on Hugging Face under Apache 2.0, allowing downloads, fine-tuning and commercial use.

The model debuted last week and scored 31 on Artificial Analysis, reportedly the highest score among open-source models outside China. It has 594 billion total parameters and activates 17 billion per inference. It supports image input and up to one million Token units of context.

Despite its name, the hardware requirements are substantial. The 2 official weight versions are NVFP4, which needs at least 500GB of GPU memory, equivalent to 2 NVIDIA B300 or 4 H200 cards, and MXFP8, which needs at least 800GB and scores just 1 to 2 percentage points higher.

The model supports vLLM, SGLang and Unsloth and is available by API through NVIDIA, Fireworks, Together and Cloudflare.