According to CoinMeta, Thinking Machines, previously founded by OpenAI CTO Mira Murati, has officially released the weights of two models, Inkling and Small. These models are now available for download, fine-tuning, and commercial use. The model was released last week and scored 31 points on a comprehensive evaluation, making it the highest-scoring open-source model outside of China at present. It boasts 594 billion parameters, activates 17 billion of these parameters per inference, and supports image input as well as up to 1 million token contexts. The official provides two versions of the model weights: the nvfp4 version requires 500GB of video memory, which is equivalent to using 2 NVIDIA B300 GPUs or 4 H200mxfp8 GPUs; the GPU version requires 800GB of video memory and only scores 1 to 2 percentage points higher in evaluations. The model is compatible with inference frameworks such as VLLM, SGLang, and Unsloth, and can be called through platforms like NVIDIA, Fireworks, Together, and Cloudflare.