Former OpenAI CTO does what Altman won't: releases a frontier AI model that's actually open

If you’re in the market for a frontier-class open weights model, your options are few and far between outside of the Chinese model houses. With the Wednesday release of a new model code-named "Inkling", an outfit called Thinking Machines Lab aims to change that. Founded in early 2025 by former OpenAI CTO Mira Murati, Thinking Machines' first model is a big one. Weighing in at 975 billion parameters, the model requires more than two terabytes of GPU memory — a quantity present in around eight of Nvidia's B300 accelerators, or sixteen H200s — to run at its native 16-bit precision. If that’s asking too much of your hardware, Thinking Machines has also released a NVFP4 quantized version of the model capable of running on half the GPUs. This makes it the largest American open weights model to date, and comparable to Chinese models like DeepSeek V4, GLM 5.2, and Kimi K2.6 in terms of size and capabilities. Take these claims with a grain of salt — gaming AI benchmarks isn’t exactly difficult — but Thinking Machines says Inkling is competitive with these models in a variety of workloads, although its benchmark charts also show it trailing proprietary models like Anthropic’s Claude and OpenAI’s GPT. Thinking Machines describes the model as being highly adaptable, intended for use by developers building AI apps, but suitable for general purpose applications like chat bots. And because it’s being released under a highly permissive Apache 2.0 license, end users are free to fine tune it for their specific use case. The company's Tinker platform offers tools to do just that. In fact, Thinking Machines boasts that the model is capable of writing its own fine tuning scripts to refine its behavior, teach itself new skills, and evaluate its abilities. Other notable features include support for a million-token context, which you can think of as the model’s short-term memory. This should help it wrangle large code bases and needle-in-the-haystack type search problems. While Thinking Machines admits the model’s mixture of experts (MoE) architecture was inspired by DeepSeek-V3, the company says it trained Inkling from scratch using Nvidia GB300 NVL72 systems and 45 trillion tokens worth of text, images, audio, and video. In total, the model features 256 routed exports and two shared ones. The model generates each token by six experts, totaling about 41 billion parameters. So, in spite of its size, the model should be able to churn out tokens at about the same rate as DeepSeek V4 when running on the same hardware. Like most LLMs today, Inkling is a “reasoning model” which is to say it’s been trained using reinforcement learning (RL) to use chain of thought to “think” through requests before responding. The model developer claims to have tuned the model to use these thinking tokens more efficiently and that Inkling therefore matches Nvidia’s Nemotron 3 Ultra, up to now the largest and most capable American open weights model out there at 550 billion parameters, on Terminal Bench 2.1 using roughly a third the tokens. Thinking tokens may make models more capable and less likely to hallucinate, but the capability comes at a cost. Those tokens are billed like any other and so the longer the model thinks, the larger users' bills become. Speaking of APIs, Inkling is available starting today on Thinking Machines’ Tinker platform, which in addition to model access also offers tools for customization and fine tuning. The company is also working to bring the model to 3rd-party API services including TogetherAI, Fireworks, Modal, Databricks, and Baseten. If you prefer to evaluate the model on your own hardware, it’s available for download on popular model repos like Hugging Face. At launch, the model claims support for a broad range of inference engines including vLLM, SGLang, Miles, TokenSpeed, and Llama.cpp. Inkling is the first of several new models under development by Thinking Machines. Alongside its flagship model, the company is also previewing Inkling-Small, a 276-billion-parameter MoE model with 12 billion active parameters for those prioritizing latency over throughput and quality. Thinking Machines — which shares its name with the fictional supercomputer maker immortalized in 1993's Jurassic Park — is currently in the process of finalizing the model and plans to release its weights once testing is complete. ®