Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, released its first AI model on July 15, 2026 — an open-weight Mixture-of-Experts system called Inkling, with 975 billion total parameters and native support for text, image, audio, and video. It is the company's first concrete public proof point after more than a year of building largely out of sight.
Key takeaways
- Inkling: MoE with 975B total parameters, ~41B active per task, trained on 45 trillion tokens
- Open-weight — downloadable and fine-tunable by anyone
- The company explicitly states Inkling is not the best available model — it is designed for well-rounded, customizable performance
- Revenue model: platform Tinker (fine-tuning, hosting) — not model API access
- Efficiency claim: 3x fewer tokens than Nvidia Nemotron 3 Ultra at equivalent coding performance
Architecture and what Inkling can do
Inkling is a Mixture-of-Experts model — an architecture where the model contains many specialized subnetworks but activates only a subset for each task. Total parameter count is 975B, but only around 41B are active during any given inference pass. This is the standard trade-off: large enough to be broadly capable, but faster and cheaper to run than a dense model of comparable total scale. Thinking Machines Lab built this on 45 trillion tokens of text, image, audio, and video.
According to the company's own benchmarks, Inkling uses three times fewer tokens than NVIDIA Nemotron 3 Ultra to reach equivalent coding performance.
Open-weight strategy and the Tinker platform
Inkling is open-weight?open-weight: a model whose parameters (weights) are publicly released for download — anyone can run or fine-tune it without an API license. OpenAI, , and Google sell API access to closed models. Thinking Machines bets on a different model: the model as a starting point, with organizations building differentiated solutions on top. Revenue is expected from Tinker: a platform for fine-tuning, adapting, and hosting models built on top of Inkling.
The company used some external open-weight models — including Moonshot AI Kimi K2.5 — to generate early post-training data before reinforcement learning took over. The next model will use a fully self-contained pipeline.
The market context behind Thinking Machines argument
Hugging Face CEO Clement Delangue predicts frontier models will be reserved for experimentation while production AI shifts to open alternatives. , Microsoft CEO, warned on July 13 that using closed external AI models hands those vendors institutional knowledge encoded in thousands of prompts.
Why it matters
Inkling is not stronger than Claude Opus or GPT-5.6 — the company says this directly. But it is the first large open model explicitly designed as a starting point for fine-tuning rather than a finished product. If Inkling plus Tinker deliver meaningfully better outcomes for specialized deployments at lower cost, Thinking Machines can capture a real enterprise position.
What is next
- Thinking Machines stated the next model will not use post-training data from external models — the full pipeline will be internal
- Tinker is available, but enterprise pricing terms have not been publicly announced
- Nvidia confirmed a significant investment at the March 2026 partnership announcement — the external funding round discussed in fall 2025 remains unconfirmed
