In February 2025, Mira Murati walked away from OpenAI. She had been the company's Chief Technology Officer. She helped build the most commercially successful closed-source AI in history. She left with a cohort of senior OpenAI researchers. Within months, Thinking Machines Lab raised a record $2 billion seed round at a $12 billion valuation. Backers included Nvidia, AMD, Cisco, Andreessen Horowitz, and Jane Street.
On July 15, 2026, Thinking Machines released its first model: Inkling.
The specs are impressive. Inkling is a mixture-of-experts transformer with 975 billion total parameters and 41 billion active per token. It was pretrained on 45 trillion tokens spanning text, images, audio, and video. It supports a 1-million-token context window. It is released under an Apache 2.0 license, with full weights available for download on Hugging Face.
But the most remarkable thing about Inkling is not what it is. It's what it acknowledges.
The Architecture That Tells the Story
Thinking Machines did not hide where Inkling came from. In its own published architecture, the company says Inkling's design "largely follows DeepSeek-V3." That is a Chinese model. Inkling also adopts DeepSeek's auxiliary-loss-free load balancing and the same MoE design principles.
The Chinese influence did not stop at architecture. Inkling's post-training started with supervised fine-tuning on synthetic data. That data came from open-weight models including Kimi K2.5, a Chinese model from Moonshot AI. The Financial Times reports Inkling's architecture was built on DeepSeek-V3. It was improved after initial training with data from Kimi K2.5.
A US lab staffed by the people who built ChatGPT adopted a Chinese architecture. It used a Chinese model to train its flagship release.
Forbes put it bluntly in its headline: "Murati Knows OpenAI's Secrets. Her AI Suggests She Prefers China's." The Financial Times told a similar story. Its headline: "Mira Murati's Thinking Machines draws from Chinese rivals in debut AI model."
The Double Standard Nobody Wants to Address
For eighteen months, Washington has accused Chinese labs of stealing American intellectual property. DeepSeek stunned the AI community in early 2025. US officials immediately claimed DeepSeek had distilled OpenAI's models. Congressional investigators alleged that Chinese AI firms ran coordinated distillation campaigns against US frontier models. The word used in Washington was theft.
Now an American lab has adopted a Chinese architecture. It distilled a Chinese model to train its own flagship release.
Nobody is calling this theft. And nobody should. The Chinese models in question are open-weight and permissively licensed. Learning from published work is how science advances.
But the asymmetry in rhetoric is now impossible to ignore. When Chinese labs learn from US models, it is theft. When US labs learn from Chinese models, it is engineering. When China builds competitive AI, it is a national security threat. When the US builds AI on Chinese foundations, it is innovation.
Forbes captured the contradiction. Its quote: "This highlights a US double standard: Chinese use of US models is 'theft,' but US adoption of Chinese open models is 'engineering'."
The Performance Reality
Murati's team has been unusually honest about Inkling's capabilities. The company stated explicitly that Inkling is "not the strongest model available today, open or closed."
On benchmarks like Humanity's Last Exam, Terminal Bench, and SWE-Bench Verified, Inkling trails the top Chinese open models. That includes Zhipu's GLM 5.2 and Moonshot's Kimi K2.6. Artificial Analysis, a third-party tester, gave Inkling a composite score of 41. That beats Nvidia's Nemotron 3 Ultra and Google's Gemma 4 31B. It still trails the top Chinese models.
The smaller sibling, Inkling-Small, has 276 billion total parameters and 12 billion active. On agentic benchmarks, it matches models four times its size. SWE-Bench Verified comes in at 80.2 percent. Terminal Bench 2.1 scores 64.7 percent. AIME 2026 reaches 95.1 percent. These are not weak numbers. They are simply not the best numbers.
But Thinking Machines is not trying to win leaderboards. The company explains its goal: "Open-source models' primary value does not lie in pushing the performance ceiling, but in finding a balance between capability, inference cost, native multimodality, and fine-tunability." The model is designed to be a base for enterprise customization, not a benchmark champion.
The Pricing Strategy
The pricing tells the strategic story. On Thinking Machines' Tinker API, serverless inference costs $0.30 per million input tokens. Output runs $1.20 per million tokens at 256K context. That is roughly half the cost of OpenAI's Luna ($0.20/$1.20) and well below Kimi K3's $3.00/$15.00.
DeepSeek's V4-Flash-0731, at $0.14/$0.28, remains cheaper. But Inkling-Small offers what DeepSeek does not. It packs a 1-million-token context window and native multimodal support into one open-weights package.
The two-model strategy mirrors what DeepSeek executed with Flash and Pro. Inkling handles the highest-complexity reasoning tasks. Inkling-Small takes agentic and efficiency-sensitive workloads.
The Geopolitical Dimension
The US-China dimension is unavoidable. DeepSeek and Moonshot now ship capable open-weights models at aggressive prices. The narrative has been that US labs are losing the open-weights race. Inkling does not end that narrative.
But it does complicate it. A US lab built its first model on Chinese architecture and Chinese training data. It is staffed by OpenAI alumni and backed by America's top venture firms. It is competing on price and customizability, not raw performance. It is playing China's game — and openly acknowledging it.
Forbes warns this shift "risks structurally disadvantaging US firms, forcing them to build on weaker foundations while global competitors use the best available."
The center of gravity for open AI, Forbes noted, is shifting to China. The most remarkable part of that shift? An American lab, founded by the people who built the most successful closed-source AI in history, now leads part of it.
Sources
- Forbes (July 15, 2026)
- Financial Times (July 16, 2026)
- Bloomberg (July 15, 2026)
- TechCrunch (July 15, 2026)
- Forkast (August 1, 2026)
- 36Kr (July 16, 2026)
- Zhidongxi (Chinese tech outlet, July 16, 2026)
- Wallstreetcn (Chinese financial outlet, July 16, 2026)
Disclaimer: The analysis above is based on publicly available data as of 2026-08-03. All benchmark scores, pricing, and performance claims are sourced from the respective companies' published materials. I am not affiliated with any of the companies mentioned unless explicitly stated. For the most current information, please visit the official sources linked throughout this article.
Limitations: This analysis relies on public information available as of August 2026. Benchmark scores come from each company's published materials and third-party testers. Scores and prices change quickly in this market. A model that leads today may not lead next quarter. The quotes from Forbes and the Financial Times are reported secondhand, not from the full paywalled articles. The geopolitical framing is opinion, not legal or policy advice. Verify current numbers before making any purchase or investment decision.