Thinking Machines Lab, created by Mira Murati, former Chief Technical Officer of OpenAI, released the company ' s first self-study AI model, Inkling. Unlike the flagship models of OpenAI, Anthropic and Google, this model uses open source weights, which can be downloaded and modified directly by developers and businesses.

The launch was the first time that the company had publicly showcased its core product after about a year and a half of low-profile construction, making more specific the idea that its “business should own and customize its own AI model”.

Masterplay enterprise customization

Inkling uses a hybrid expert structure with a total parameter size of 97.5 billion, but only about 41 billion parameters are called on for a single mission. This design is common to large models and is intended to reduce operational costs and increase response speed.

According to corporate disclosures, Inkling used 45 trillion token of text, image, audio and video data training to support cross-modular reasoning. Thinking Machines did not position it as the strongest model at present, but rather emphasized its combined performance and adjustability.

According to the company, users can adjust the “think strength” of the model to the needs of speed and effect. In a programming baseline test, the number of tokens needed when Inkling had reached a level close to that of Nvidia Nemotron 3 Ultra is about one third.

Don't take the universal chat robot route.

Thinking Machines is now more like using Inkling as a starting point for second training for companies than a finished chat robot for the public. The company expects its clients to fine-tune and deploy the model through its customized platform, Tinker.

This is different from ChatGPT, Claude, Gemini. The latter enters the market in the form of a general assistant, then gradually superimposes agency and automation functions, while Thinking Machines pledges that intra-enterprise knowledge and business processes are more suitable for carrying through privatization or open models.

TechCrunch mentioned in the report that this view is gaining more support. Microsoft CEO Satya Nadella has also indicated in recent days that the use of proprietary models by enterprises tends to cover both the costs of subscriptions and the potential for their business knowledge to be absorbed by model providers in long-term interactions.

Cases, calculations and commercialization remain a concern

It was reported that Bridgewater Associates and Thinking Machines researchers had continued their training on the basis of existing open source models, combining bridge water with their own financial knowledge. The results published by both parties showed that the model achieved 84.7 per cent in the financial reasoning test, outperforming some of the head closed-source models, with operating costs of about a quarter of the latter. However, this result was derived from a self-assessment by both parties and was not an independent test.

In terms of training modalities, Thinking Machines states that the pre-training of Inkling was completed from scratch but that during the early post-training phase, some data were generated using other open-source weight models, including Kimi K2.5 of Moonshot AI. According to the company, the next generation model will be replaced by a fully owned post-training process.

On the computational side,Thinking Machines entered into a strategic collaboration with Nvidia in March this year to deploy 1 Giwa Vera Rubin computing capability, and stated that Inkling was based solely on Nvidia GB300 NVL72 training. However, companies have not yet elaborated on the income model and cost balance.

According to reports, Thinking Machines currently has about 200 employees. For this company, the publication of Inkling is not only the first open product node, but also the validation of a commercial route that differs from a head-closed-source laboratory: model weights are open, and more income comes from training, fine-tuning and hosting ecology than simply selling calls.