Tech • IA • Crypto
Mira Murati's First AI Model Is Built on China's Blueprint... Wild
9.4/10TL;DR
Thinking Machines Lab has launched Inkling, a massive open-weight AI model that prioritizes efficiency, multimodal breadth, and platform strategy over benchmark dominance.
KEY POINTS
A high-profile debut from a well-funded startup
Thinking Machines Lab, founded in 2025 by former OpenAI CTO Mira Murati, raised $2 billion at a $12 billion valuation before releasing any model. Backed by Nvidia, AMD, Cisco, Andreessen Horowitz, and Jane Street, the company initially offered only a fine-tuning API before unveiling its first full model.
Massive architecture with efficient activation
Inkling is a mixture-of-experts transformer with 975 billion parameters, but only about 41 billion activate per query, keeping costs manageable. It supports a 1 million token context window and was trained on 45 trillion tokens spanning text, images, audio, and video, enabling native multimodal reasoning.
Fully open weights under permissive licensing
The model is released under Apache 2.0, with complete weights available publicly, including optimized versions for Nvidia Blackwell hardware. Developers can run, fine-tune, and deploy it freely, positioning Inkling as one of the largest open-weight AI systems available.
Performance trails leading competitors
Inkling does not lead benchmarks. It scores 29.7% on Humanity’s Last Exam, compared with 40.1% for GLM 5.2 and 53.3% for Claude Opus 5. On SWE-bench Verified, it achieves 77.6%, and 63.8 on Terminal Bench, both behind top models, including several Chinese open systems.
Strength in efficiency and real-world workflows
The model is designed as a balanced generalist rather than a leaderboard specialist. It matches competitors like Nvidia NeMoTron 3 Ultra on some tasks while using roughly one-third of the tokens, significantly lowering operational costs in large-scale deployments.
Advanced agentic and multimodal capabilities
Inkling can transcribe audio, analyze images, process charts, and execute tool calls such as Python-based image manipulation mid-reasoning. Demonstrations include generating full web apps, refining games through iterative feedback, and producing structured documents like multi-page reports from a single prompt.
Uncertainty-aware reasoning (“epistemics”)
A distinguishing feature is its ability to express uncertainty. Trained with probabilistic scoring methods, the model assigns confidence levels and abstains when evidence is weak. On Forecast Bench, it achieves a Brier score of 61.1, outperforming some proprietary systems and matching Gemini 3.1 Pro.
Rigorous verification and safety training
Training included dual automated graders: one evaluating answer quality and another verifying factual claims via web search. Safety testing covered cyber risks, manipulation, and weapons-related prompts. On the Fortress benchmark, Inkling scored 78% on adversarial refusal while maintaining 95.9% accuracy on benign queries.
Built using global open AI influences
The architecture closely follows DeepSeek V3, and training incorporated synthetic data from models like Kimi K2.5. This reflects a broader industry trend of cross-pollination between U.S. and Chinese open models, challenging narratives around intellectual property and innovation flows.
Platform strategy over model supremacy
The company’s business model centers on Tinker, its paid fine-tuning platform, rather than charging for API usage. Inkling serves as a customizable base. In one case, Bridgewater Associates fine-tuned a model achieving 84.7% on financial benchmarks at under 10% of the cost of leading proprietary systems.
Geopolitical and market implications
U.S. restrictions on Chinese AI models are reshaping adoption. With firms previously routing up to 45% of enterprise AI usage through Chinese systems, regulatory pressure is pushing demand for Western open alternatives. Inkling is positioned to fill that gap as a “permitted” option.
Emergence of a smaller competitive variant
A preview of Inkling Small shows strong results despite fewer active parameters (12 billion), outperforming the larger model on several benchmarks, including 88.3% vs. 87.2% on GPQA Diamond, suggesting efficiency gains may outpace raw scale.
CONCLUSION
Inkling signals a shift in AI competition from raw model performance to efficiency, openness, and platform ecosystems, with strategic and geopolitical factors increasingly shaping which technologies gain adoption.