Ex-OpenAI CTO Murati’s Thinking Machines Drops Inkling: A 975B Parameter Model That Leads US Labs but Trails China
Former OpenAI CTO Mira Murati’s new startup, Thinking Machines, has released Inkling, a 975-billion-parameter AI model. The model outperforms leading US labs on key benchmarks but falls short of top Chinese rivals. Inkling represents a major push to reclaim US leadership in frontier AI, though it highlights the rapidly closing gap.
Inkling’s Performance Edge
Inkling achieves state-of-the-art results across multiple reasoning and language tasks. The 975B parameter model beats GPT-4, Claude 3.5, and Gemini Ultra on several benchmarks, including math, coding, and scientific reasoning.
- Mathematical reasoning: Inkling scores 90.1% on GSM8K, exceeding GPT-4’s 87.3%.
- Code generation: It outperforms on HumanEval and MBPP, with a 92.5% pass rate.
- Scientific knowledge: On MMLU, Inkling achieves 89.7%, surpassing previous US leaders.
“Inkling shows that US labs can still push the frontier, but we cannot be complacent,” said a Thinking Machines researcher. “The gap with China’s best models is now measured in months, not years.”
The China Factor
Despite leading US labs, Inkling trails Chinese models like DeepSeek-V3 and Qwen2.5-72B. On the difficult MATH-500 benchmark, Chinese models achieve 94.2%, while Inkling scores 91.8%. On the Chinese-language benchmark C-Eval, Inkling lags by 3.5 points.
- Parameter efficiency: Chinese models often use fewer parameters but achieve comparable or better results, suggesting architectural advantages.
- Training data diversity: Chinese models benefit from large-scale, high-quality multilingual data, including more recent web content.
- Hardware constraints: US export controls on advanced chips have forced Chinese labs to innovate with less compute, leading to more efficient training methods.
Why Inkling Matters
Inkling is a signal that the US still has top-tier talent and infrastructure. Murati’s team, drawn from OpenAI, DeepMind, and Meta, assembled a massive dataset and trained Inkling on a cluster of 100,000 H100 GPUs. The model is designed for both research and commercial use.
- Open-weight release: Thinking Machines is releasing Inkling’s weights under a research license, allowing academics and startups to build on it.
- Safety emphasis: The company claims to have implemented new alignment techniques, including red-team testing and constitutional AI guardrails.
- Cost implications: Training Inkling cost an estimated $500 million, highlighting the increasing capital required to compete.
The Bigger Picture
The US-China AI race is now neck-and-neck on raw performance, but China leads in speed of iteration. While Inkling represents a technical achievement, the gap is narrowing fast. Chinese labs have released dozens of models over the past year, each improving on the last. The US, by contrast, has seen fewer major releases.
- Government policy: US export controls have slowed China’s access to cutting-edge chips, but Chinese firms have developed alternative hardware and software workarounds.
- Open-source ecosystem: Chinese models are often open-source, accelerating global adoption and improvement. Inkling’s open-weight release aims to counter that.
- Future outlook: If the trend continues, China could overtake the US in frontier AI within 12 to 18 months, according to some analysts.
What’s Next for Thinking Machines
Murati’s startup plans to release smaller, more efficient versions of Inkling for edge devices and specific industries. The company is also exploring partnerships with cloud providers to offer API access. The long-term goal is to build a general-purpose AI that can be deployed safely and at scale.
“We are not just building a model; we are building the infrastructure for responsible AI,” Murati said in a recent interview. “Inkling is a step toward that vision.”
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.