America's first open-source model officially launches: Inkling 2 Small goes open source, requiring at least 500GB of VRAM to run.

Beating AI News Flash: Thinking Machines, founded by former OpenAI CTO Mira Murati, has officially released the model weights for Inkling 2 Small. The model is now available on Hugging Face under the Apache 2.0 license, allowing downloads, fine-tuning, and commercial use.

Inkling 2 Small was released last week and achieved a composite score of 31 on Artificial Analysis, making it the highest-scoring open-source model outside China. It has 594 billion total parameters, activates 17 billion per inference, supports image input, and offers a context window of up to 1 million tokens.

Despite being called Small, the hardware requirements are quite high. The official release provides two weight versions: the NVFP4 version requires at least 500GB of GPU memory, equivalent to 2 NVIDIA B300s or 4 H200s; the MXFP8 version requires at least 800GB of GPU memory, with benchmark scores only 1 to 2 percentage points higher than the former.

Currently, the model supports inference frameworks such as vLLM, SGLang, and Unsloth, and can also be accessed via APIs through platforms including NVIDIA, Fireworks, Together, and Cloudflare.