Mira Murati‘s AI startup, Thinking Machines Lab, released the Inkling AI model on Wednesday, an open-weights system with 975 billion parameters that the company says rivals the top Chinese frontier models. Murati founded Thinking Machines in early 2025 after leaving her post as chief technology officer at OpenAI, and the company built the Inkling AI model under an Apache 2.0 license, letting anyone download, modify, and redeploy it without the restrictions that OpenAI and Anthropic place on their own flagship systems.
The release makes Inkling the largest American open-weights model to date, a category that until now had trailed far behind the Chinese model houses. Running Inkling at its native 16-bit precision needs more than two terabytes of GPU memory, roughly eight of Nvidia’s B300 accelerators or sixteen H200s. Thinking Machines also shipped an NVFP4 quantized version that runs on half that hardware, a concession to developers who cannot afford a full GPU cluster.
How the Inkling AI model compares to DeepSeek and Nemotron
Thinking Machines says Inkling is comparable in size and capability to DeepSeek V4, GLM 5.2, and Kimi K2.6, though its own benchmark charts show it trailing proprietary systems like Anthropic’s Claude and OpenAI’s GPT. The company also claims Inkling matches Nemotron 3 Ultra, Nvidia’s 550-billion-parameter model that had been the largest American open-weights system before this release, on the Terminal Bench 2.1 benchmark while using roughly a third the tokens. Benchmark comparisons between model makers are easy to game, so those figures are worth treating with some caution until independent testing catches up.
Architecture and training
Thinking Machines says it trained the Inkling AI model from scratch on Nvidia GB300 NVL72 systems, feeding it 45 trillion tokens of text, images, audio, and video. The model uses a mixture-of-experts design that the company acknowledges was inspired by DeepSeek-V3, with 256 routed experts and two shared ones. Each token passes through six experts, activating about 41 billion parameters at a time, which should let Inkling generate tokens at roughly the same speed as DeepSeek V4 on comparable hardware despite its larger overall size.
Like most current large language models, Inkling is a reasoning model trained with reinforcement learning to work through a chain of thought before answering. That approach can reduce hallucinations, but the thinking tokens it generates are billed the same as any other output, so longer reasoning chains raise the cost of running the model.
Access and what comes next
The Inkling AI model is available now through Thinking Machines’ Tinker platform, which also offers fine-tuning tools, and the company says it is working with third-party API providers including TogetherAI, Fireworks, Modal, Databricks, and Baseten to bring the model to their services. Developers who want to run it themselves can download the weights from Hugging Face, and the model supports inference engines including vLLM, SGLang, Miles, TokenSpeed, and Llama.cpp at launch.
Thinking Machines is also previewing Inkling-Small, a 276-billion-parameter mixture-of-experts model with 12 billion active parameters, aimed at developers who prioritize response speed over raw capability. The company says it is still finalizing that smaller model and plans to release its weights once testing wraps up.