Thinking Machines has launched Inkling, a large open multimodal language model featuring approximately one trillion parameters and a one-million-token context window. Trained on 45 trillion tokens of text, images, audio, and video, Inkling is designed to reason natively across multiple modalities. The decoder-only Mixture-of-Experts model contains 975 billion total parameters and 41 billion active parameters, utilizing relative attention, hybrid attention layers, short 1D convolutions, and a shared experts sink. Thinking Machines also released Inkling-Small, which has 276 billion total parameters and 12 billion active parameters while sharing the same underlying architecture. Inkling is available in full BF16 and NVFP4 variants, and includes speculative MTP layers for faster inference. Day-0 support is provided across major inference engines and frameworks, including transformers, SGLang, vLLM, and llama.cpp. Developers can run the model locally using quantized GGUF versions or deploy it in clusters via distribution tools like SLURM. Serverless inference and evaluation options are also accessible through Hugging Face Inference Providers, expanding accessibility for multimodal reasoning applications.
Welcome Inkling by Thinking Machines
Thinking Machines has released Inkling, a large open multimodal language model featuring approximately one trillion parameters and a one-million-token context window. The decoder-only Mixture-of-Experts model natively accepts image, text, audio, and video inputs, and is trained on 45 trillion tokens. Alongside the main model, Thinking Machines launched Inkling-Small, which has 276 billion total parameters and 12 billion active parameters.
Why it matters
The release provides the AI industry with a massive, natively multimodal open model capable of handling diverse input modalities and long contexts. Its widespread day-0 framework support enables researchers and developers to deploy and fine-tune large-scale reasoning applications immediately.