Former OpenAI CTO Mira Murati Releases Inkling, a 975B-Parameter Open Weights Frontier Model
Key Takeaways
- ▸Thinking Machines Lab released Inkling, the largest American open weights frontier model at 975B parameters, directly challenging the dominance of proprietary models
- ▸Founded by former OpenAI CTO Mira Murati, the company positions itself as releasing what OpenAI won't—a truly open frontier model under Apache 2.0 license
- ▸Inkling matches or competes with leading Chinese models (DeepSeek V4, GLM 5.2, Kimi K2.6) while offering US-based development and the flexibility for custom fine-tuning
Summary
Mira Murati, the former Chief Technology Officer of OpenAI, has founded Thinking Machines Lab and released Inkling, a 975 billion-parameter open weights model that stands as the largest American frontier-class model available to the public. The model, released under a permissive Apache 2.0 license, is positioned as a direct alternative to proprietary models from OpenAI and Anthropic, while also competing with Chinese frontier models like DeepSeek V4, GLM 5.2, and Kimi K2.6 in terms of scale and capability.
Inkling features a mixture of experts architecture with 256 routed experts, requires 8 Nvidia B300 accelerators or 16 H200s to run at native precision, and supports a million-token context window for handling large codebases and complex reasoning tasks. The model was trained from scratch on 45 trillion tokens of diverse data—including text, images, audio, and video—using Nvidia's GB300 NVL72 systems. It incorporates chain-of-thought reasoning capabilities and a reasoning token system designed to be more efficient than competitors, matching performance of Nvidia's Nemotron 3 Ultra using roughly one-third of the thinking tokens.
Thinking Machines is making Inkling available immediately through its Tinker platform, which provides tools for fine-tuning and customization, with plans to launch on third-party API services including TogetherAI, Fireworks, Modal, Databricks, and Baseten. The open release marks a significant shift in the U.S. AI landscape, as major players like OpenAI have consistently resisted releasing frontier-class open weights models.
- The model features a 1 million-token context window, efficient reasoning tokens, and multimodal training on text, images, audio, and video
- Available immediately on Thinking Machines' Tinker platform with rapid rollout planned for third-party API services including TogetherAI and Fireworks
Editorial Opinion
Inkling's release represents a watershed moment for open AI development in the United States. By departing OpenAI to build genuinely open frontier models, Mira Murati is filling a critical gap left by the industry's leading companies, which have consistently chosen to keep their most powerful models proprietary. This move may finally give American developers and researchers competitive alternatives to Chinese open models while preserving the ability to customize and fine-tune at the frontier scale—a capability that has been largely locked behind closed corporate walls.
