Thinking Machines Releases Inkling 2 Small Open-Weight Model, Claims Western Lead
The start-up founded by Mira Murati, Thinking Machines, has unveiled Inkling 2 Small, an open-weight model that the company says is the most capable of its kind from the West. The trade-off is significantly higher token consumption.
The start-up Thinking Machines has released Inkling 2 Small, a new language model available as an open-weight model. According to the company, it is the strongest Western model in this category. Thinking Machines was founded by Mira Murati, who previously served as Chief Technology Officer at OpenAI. The release fits into a phase in which developers and companies increasingly seek models whose weights they can deploy and examine themselves.
Inkling 2 Small is based on a Mixture-of-Experts architecture and has 594 billion parameters in total. With this approach, not all parts of the network are activated for every request; only certain specialized expert modules are used. The overall size of the model is therefore considerable, even though only a portion of it is used for each computation. For users, this means that the theoretical capability of the architecture cannot be equated one-to-one with the computing cost of operation.
The company points to significant progress in agentic tasks, meaning applications in which an AI system independently plans and executes multi-step processes. Such tasks are considered one of the most demanding areas for language models, because errors in an early step can propagate through the entire chain. At the same time, the reporting points to a key drawback: token consumption rises sharply with these performance gains. Since costs are typically calculated based on the tokens processed, this higher consumption is the most important economic criticism for users.
The report from The Decoder relies primarily on information provided by the manufacturer. No independent benchmarks are cited, so the claim that Inkling 2 Small is the strongest Western open-weight model has not yet been independently verified. Nor can the available information quantify how large the additional token consumption is in concrete terms or what cost implications arise in individual cases. Without such measurements, it remains open whether the performance lead justifies the additional overhead.
Open-weight models are of interest to developers and companies because the model's weights are available, allowing the system to be run locally or on one's own infrastructure. This opens possibilities for customization and control over data, but also requires dedicated computing capacity. Whether Inkling 2 Small maintains its lead in practice will therefore likely only become clear once independent tests are available and users can weigh the effort against the benefits.
Source: the-decoder.com
4
