Mira Murati, the former CTO of OpenAI, has finally revealed what she has been working on for the past eighteen months. Her debut model, Inkling, is a massive 975-billion-parameter open-weights powerhouse with a 1-million-token context window. However, a peek under the hood reveals an ironic twist: the foundation of America’s strongest open model is built on Chinese data—specifically Moonshot AI’s Kimi K2.5. Once again, technological isolationism has surrendered to pragmatism.
In this industry, we are reaching a moment of truth: high-quality data carries no passport. While Washington politicians draw lines in the sand, Thinking Machines Lab quietly utilized Chinese "intelligence" for the supervised fine-tuning (SFT) stage. This isn't just a quirk; it is a frank admission that creating a top-tier product today is impossible without a synthetic symbiosis. The developers openly acknowledge using Kimi, though they clarify that the primary breakthrough occurred later during reinforcement learning (RL). Nevertheless, the fact that American pride is kneaded with a competitor's code serves as an elegant mockery of the idea of closed digital borders.
Technical Honesty Over the Numbers Race
Inkling’s architecture utilizes a Mixture of Experts (MoE) approach that looks suspiciously like DeepSeek-V3’s design: 256 routed experts and two shared ones. For every token, 41 billion parameters are activated. Eschewing traditional encoders, the team opted for a single transformer for all modalities. Images are processed as patches and audio via spectrograms, all digested by shared weights.
"Inkling is not the highest-performing model available today, whether closed or open," Thinking Machines admits, shifting the focus from benchmarks to epistemology and confidence calibration.
Instead of chasing leaderboard peaks at any cost, the model was trained to recognize the limits of its own knowledge. During RL training, Inkling was drilled on questions with known outcomes to reward predictive accuracy. On the ForecastBench test, the model matched the performance of Gemini 3.1 Pro and Grok 4.3. For business leaders, this is more critical than abstract scores: a model capable of self-doubt is less likely to hallucinate during mission-critical processes.
The Economics of "Thinking Effort"
Inkling’s standout practical feature is its adjustable "thinking effort." Unlike competitors with rigid presets, users can manually set a reasoning intensity slider from 0.2 to 0.99. This is a direct response to enterprise demands for cost optimization. On the Terminal Bench 2.1 coding benchmark, the model matched Nvidia’s Nemotron 3 Ultra in quality while consuming three times fewer tokens. Why pay for deep contemplation when basic logic suffices?
During RL training, the model began optimizing its internal monologue autonomously. Much like a stenographer, it started stripping away articles and conjunctions, adopting a telegraphic style without losing quality. This supports the hypothesis that AI pursues its own linguistic efficiency when forced to conserve resources. The Tinker customization platform even looks like a harbinger of the end of traditional development: in one demo, Inkling wrote the code for its own fine-tuning, generated the data, and updated its own weights.
Synthetics as the New Standard
While Inkling trails top Chinese models like Kimi K2.6 in raw power, it currently stands as the favorite among U.S. open systems, outperforming Nemotron 3 Ultra. Its top-ten position in the Design Arena alongside Claude Opus 4.6 confirms that the era of open-source AI being the "poor relative" is over. However, reliance on foreign synthetic data creates risks. If the entire market begins training on one another’s outputs, we face intellectual stagnation—a feedback loop where one model's errors are indefinitely replicated across the weights of others.
Murati’s strategy is clear: occupy the high-performance open-model niche effectively vacated by Meta. Backed by Nvidia investment and a Google Cloud deal, Thinking Machines is building a world where the model is just the core, and the real value lies in owning your own "intellectual factory." You can download Inkling’s weights from Hugging Face and see for yourself: its resistance to censorship and honest calibration allow for AI deployment without the fear of an unexpected corporate scandal.