Inkling: Mira Murati's First AI Model Is Open-Weight
July 19, 2026
Mira Murati's Thinking Machines Lab shipped its first model this week. It skipped the API business entirely.
Inkling is open-weight — download it, inspect it, fine-tune it, self-host it. No per-token billing, no rate limits, no vendor lock-in. It's also the largest US-built open-weight model on Hugging Face right now, at 975 billion parameters.
What Inkling Actually Is
Inkling is a mixture-of-experts model with 975B total parameters, but only about 41B activate per token — that's what keeps inference fast and affordable despite the huge parameter count. It was trained on 45 trillion tokens across text, image, audio, and video, and it reasons natively across all four.
- License: Apache 2.0, full weights on Hugging Face
- Context window: up to 1 million tokens; Tinker offers 64K and 256K tiers
- Adjustable "thinking effort" — trade reasoning depth for lower cost and latency
- Trained on Nvidia's latest infrastructure
Thinking Machines is also previewing a lighter Inkling-Small, with weights coming after further testing.
The Business Model Is the Real Story
Thinking Machines isn't charging for API calls. It's charging for Tinker, its paid fine-tuning platform. Give away the base model, sell the tooling that makes it useful for your specific problem — that's the pitch.
The proof point: Bridgewater Associates used Tinker to fine-tune an open model on proprietary financial data. The result scored 84.7% on financial reasoning benchmarks — beating proprietary alternatives at under 10% of the cost, according to the two companies' own evaluation.
It's Not Trying to Win Benchmarks
Thinking Machines says it directly: Inkling is "not the strongest overall model available today, open or closed." That's an unusual thing for a lab to admit on launch day.
The bet is that businesses care more about control than leaderboard position — owning the fine-tuned result, running it on your own infrastructure, and not handing your prompts and data to a closed API.
Why This Matters: China Owns the Open-Weight Race
Open-weight AI has been a mostly Chinese story for the past year — DeepSeek, Qwen, Kimi, and Meituan's LongCat-2.0 have all shipped competitive open models while Meta pulled back toward closed releases after Llama 4 underwhelmed.
Inkling gives US developers a serious open-weight alternative that isn't routed through a Chinese lab — relevant if your legal or security team has opinions about model provenance.
Key Takeaways
- Inkling is a 975B-parameter MoE model (41B active), Apache 2.0, live on Hugging Face today
- Revenue comes from Tinker fine-tuning, not API fees — the model itself is free
- It's not benchmark-leading by design — customizability is the pitch, not raw capability
- It's the largest US-built open-weight model, closing a gap Chinese labs have dominated
Sources: TechCrunch, Axios, Fortune, eWeek
Frequently Asked Questions
What is Thinking Machines' Inkling model?
Inkling is the first AI model built by Mira Murati's Thinking Machines Lab, released July 15, 2026. It's a 975-billion-parameter mixture-of-experts model, open-weight under Apache 2.0, with full weights available on Hugging Face.
How does Thinking Machines make money from a free model?
It doesn't charge for API access to Inkling itself. Revenue comes from Tinker, its paid platform for fine-tuning the model on proprietary data, offered with 64,000- and 256,000-token context tiers.
Is Inkling the best open-weight model available?
No, and Thinking Machines says so directly — Inkling is "not the strongest overall model available today, open or closed." The company is betting on customizability through Tinker rather than raw benchmark performance.
Can I run Inkling on my own infrastructure?
Yes. Because it's released under Apache 2.0 with full weights on Hugging Face, you can download, self-host, and fine-tune Inkling without routing queries through Thinking Machines' API.