Thinking Machines Lab has released Inkling, the first model from the company founded by Mira Murati (former CTO of OpenAI): 975 billion total parameters, 41 billion active in Mixture-of-Experts architecture, Apache 2.0 license, multimodal (text, image, audio as input, text as output), up to 1 million token context window. As we reported on July 17, the model had already launched; now TechCrunch adds details on the funding round and strategic vision.

Why it matters. Thinking Machines raised $300 million in pre-seed and moved the model from pretraining to release in nine months (OpenAI took five years, Anthropic three). The thesis behind Inkling: generic one-size-fits-all models sold as closed services underperform compared to models that organizations adapt themselves. Inkling doesn’t compete on general benchmarks — the company states this explicitly — but aims to be the starting point for fine-tuning on Tinker, the company’s model customization platform. Whoever adapts becomes responsible for the security of their own versions (and requires serious ML talent). The Bridgewater case is the concrete argument: an open model trained on internal financial expertise beat top proprietary models on financial reasoning at one-fourteenth the cost. The numbers come from internal evaluation, not independent benchmarking.

Distillation. Thinking Machines pre-trained Inkling from scratch on 45 trillion tokens, but used other open-weight models (including Kimi K2.5 from Moonshot AI) to generate part of the post-training data before large-scale reinforcement learning took over. It’s a form of distillation limited to post-training, disclosed in release materials.

In detail

The context: nine months versus five years.

OpenAI took roughly five years to bring its technology to market and show revenue; Anthropic about three. Thinking Machines claims to have done the same in nine months: from pretraining to public release with day-0 ecosystem support (vLLM, SGLang, Modal, Baseten, Databricks, Hugging Face, community quantization). The $300 million pre-seed round is unusual in size and stage: typically pre-seed is orders of magnitude smaller.

The model is 975B total parameters, 41B active (Mixture-of-Experts), trained on 45 trillion tokens (according to readers of release materials; one source says 48T, the majority 45T). Supports up to 1M token context in open-weight checkpoints, but the API on Tinker offers 256K (with differentiated pricing for 64K and 256K). Text-only output for now, but multimodal input (text, image, audio).

The thesis: customization beats generality.

Thinking Machines argues that AI trained centrally by a lab and then frozen underperforms compared to AI that organizations shape themselves, because much expertise is specific and lives with the people who hold it. Inkling is designed as a foundation model to fine-tune, not as an already-finished general-purpose chatbot. The company explicitly states that Inkling isn’t the strongest model available today, either open or closed. The goal is balanced performance and adaptability.

The concrete case cited is Bridgewater Associates (the world’s largest hedge fund, which is not a Thinking Machines investor): they took an existing open-source model and trained it further on internal financial expertise. Result: 84.7% on financial reasoning tests, beating top proprietary models, with inference costs about one-fourteenth of closed models. The numbers come from joint Thinking Machines–Bridgewater evaluation, not independent benchmarks.

Arguments against closed models.

The post Thinking Machines published the week before launch set the stage: centrally-trained and then-fixed AI underperforms because it doesn’t capture organizations’ specific knowledge. Others are converging on similar arguments. Satya Nadella (Microsoft CEO, billion-dollar investor in OpenAI and Anthropic) wrote Sunday that companies using proprietary models pay twice: once in subscriptions, and again by surrendering business knowledge embedded in prompts and corrections, which can end up in future model versions. Clem Delangue (Hugging Face CEO) told TechCrunch last week that frontier models will remain for experimentation and high-value tasks, while most production AI work will shift to private or open-source alternatives — exactly the split Thinking Machines is building on.

Distillation and transparency.

Thinking Machines states it pre-trained Inkling from scratch, but used other open-weight models (including Kimi K2.5 from Moonshot AI) to generate part of the initial post-training data, before large-scale reinforcement learning became the main method. It’s a form of distillation limited to post-training and disclosed in materials. Not distillation from competing proprietary models (a practice that drew scrutiny in the industry), but from other open models.

Ecosystem support and benchmarks.

Inkling launched with immediate support on vLLM, SGLang, Modal, Baseten, Databricks, Hugging Face, and community quantization tooling (thanks to @danielhanchen). Independent commentators labeled it as the strongest US open-weight model so far, but generally behind the best Chinese open-weight models and top closed models on some benchmarks. Artificial Analysis ranks it first among US open-weight models on the Intelligence Index, but behind top Chinese ones.

Inkling supports controllable reasoning effort (numerical effort levels) and up to 1M token context in checkpoints. The company states that on a coding benchmark, Inkling uses one-third the tokens of Nvidia Nemotron 3 Ultra to reach the same performance.

Implications for builders.

If the thesis holds, the business model shifts: instead of paying for API access to a generic closed model, you pay once for compute and ML talent to train an open model on your data, then inference at lower marginal cost. It requires serious internal ML capability though: fine-tuning isn’t configuration, and responsibility for the security of customized versions falls on the organization. Not everyone has those teams.

For closed labs, pressure increases: if Nadella (billion-dollar investor in OpenAI) publicly says proprietary models make you pay twice, the message to CTOs is clear. The counter-argument remains: closed models work out of the box, without assuming an ML team. The battle is over which friction matters more: paying the subscription or hiring the engineers.