Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, has released its first model, a 975-billion-parameter behemoth called Inkling that is now the largest American open-weight AI model on the market.
Released under the permissive Apache 2.0 license, Inkling is a mixture-of-experts system that activates roughly 41 billion of its total parameters for any given task. It was trained from scratch on 45 trillion tokens spanning text, images, audio, and video, and supports a 1-million-token context window. The architecture uses 256 routed experts plus two shared experts, with six experts activated per token, a design inspired by DeepSeek V3.
The model is competitive with Chinese open-weight frontier models such as DeepSeek V4, GLM 5.2, and Kimi K2.6, though its benchmark scores still trail proprietary systems from Anthropic and OpenAI. Its significance lies less in topping every chart and more in filling a gap that has frustrated the open-source AI community: no American lab has released a genuinely capable open-weight model since Meta’s Llama 3.1 in 2024.
Inkling requires substantial hardware to run at native 16-bit precision, around two terabytes of GPU memory, equivalent to roughly eight of Nvidia’s B300 accelerators or 16 H200s. Thinking Machines has also released an NVFP4 quantized version that needs half the GPUs. The company’s Tinker platform offers tools for fine-tuning and customization.
One of the model’s more unusual capabilities is its ability to write its own fine-tuning scripts. “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,” The Register reported. The company says Inkling uses chain-of-thought reasoning trained via reinforcement learning, and claims it can match Nvidia’s Nemotron 3 Ultra on the Terminal Bench 2.1 benchmark while using roughly one-third the thinking tokens, though those tokens still cost money, and longer reasoning chains mean higher API bills for users.
Inkling is available on Hugging Face under the identifier `thinkingmachines/Inkling` and through third-party API providers including TogetherAI, Fireworks, Modal, Databricks, and Baseten. Supported inference engines include vLLM, SGLang, Llama.cpp, and TokenSpeed.
A smaller variant, Inkling-Small, has been previewed as a 276-billion-parameter MoE with 12 billion active parameters, targeting latency-sensitive applications. Its weights will be released after testing completes.
For Murati and Thinking Machines Lab, which spent roughly 18 months building largely out of public view, Inkling is the first public proof point in a broader bet against the one-size-fits-all AI model philosophy that has dominated the frontier.
Sources: “Former OpenAI CTO does what Altman won’t, releases a frontier AI model that’s actually open” (The Register, July 16, 2026); “Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling” (TechCrunch, July 15, 2026)