The fact. Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has released Inkling, a 975-billion parameter frontier model — the largest American open-weights model ever released. Codenamed Inkling, the model uses a mixture-of-experts (MoE) architecture inspired by DeepSeek-V3, was trained from scratch on 45 trillion tokens (text, images, audio, and video) across Nvidia GB300 NVL72 clusters, and is distributed under the Apache 2.0 license. It supports a 1 million-token context window. The company has also released an NVFP4 quantized version requiring half the hardware.
Context. The high-performance open model landscape has been dominated by Chinese companies such as DeepSeek (V4), GLM (5.2), and Kimi (K2.6). In the US, the largest open model until now was Nvidia's Nemotron 3 Ultra at 550 billion parameters. Thinking Machines — founded in early 2025 by Murati after her departure from OpenAI in September 2024 — arrives to fill this gap with a model the company claims is competitive with Chinese leaders across various workloads. Inkling's MoE architecture uses 256 routed experts plus two shared ones, activating 6 experts (~41B parameters) per token for efficient inference.
Analysis. Inkling represents a milestone for AI transparency. While Sam Altman's OpenAI increasingly moves toward closed, proprietary models, Murati — who oversaw the launch of GPT-4 and ChatGPT — delivers a comparably-sized model under a permissive license. The Apache 2.0 choice is significant: it allows fine-tuning, redistribution, and commercial use without restrictions. The model can also write its own fine-tuning scripts, a step toward self-improving AI agents that could dramatically reduce the human effort needed for domain adaptation. However, the company's own benchmarks show Inkling still trailing proprietary models like Claude and GPT — the gap between open and closed persists, though it is narrowing. Inference costs remain high: running at native 16-bit precision requires more than two terabytes of GPU memory, demanding 8x B300 or 16x H200 Nvidia GPUs.
What to watch. Availability via the Tinker platform (with customization tools) and arrival on third-party APIs like TogetherAI, Fireworks, and Databricks will determine real-world adoption. The move reframes the AI sovereignty debate: governments and enterprises avoiding Chinese models for geopolitical reasons now have a truly open American alternative. Real-world performance (outside benchmarks) will tell whether Inkling is a legitimate frontier competitor or simply the largest — not necessarily the best — open model available. The efficiency of Inkling's thinking tokens compared to competitors like Nemotron 3 Ultra (matching performance with a third of the tokens) could be a key differentiator for cost-conscious developers.
Source: The Register
