Technical Specifications and Architecture
Thinking Machines has announced the release of Inkling 2 Small, the second generation of its model series. The new model utilizes a Mixture-of-Experts (MoE) architecture, featuring a total of 594 billion parameters, with 17 billion parameters active during inference. This represents a significant increase in scale compared to the previous Inkling Small model, which contained 276 billion total parameters and 12 billion active parameters, according to data shared on Reddit.
The model supports multimodal inputs, specifically image processing, and features “variable effort” capabilities. According to the company’s model card, the performance improvements are attributed to more efficient pre-training and a large, unified reinforcement learning (RL) run.
Performance Benchmarks
Inkling 2 Small has achieved a score of 31 on the Artificial Analysis Intelligence Index. The company reports that this score reflects advancements in several key areas, including reasoning, vision, knowledge work, and agentic coding. These gains are positioned as a substantial improvement over the capabilities of the previous model generation.
Availability and Licensing Roadmap
The deployment of Inkling 2 Small will follow a phased approach. Initially, the model will be available through the Tinker platform with closed weights. This strategy aligns with the company’s previously outlined “safe path to open weights.”
Thinking Machines has stated its intention to release the full weights under an Apache 2.0 license next week, following the conclusion of external red-teaming and safety testing. Users can currently access the model via the Tinker Playground for testing, and the platform will support fine-tuning and sampling. The company has indicated that model checkpoints will be made available for download once the weights are released.
Community and Industry Reception
The announcement has drawn varied reactions from the developer community. On Reddit, some users expressed concerns regarding the hardware requirements for running models of this scale, noting that a 594B parameter model poses significant challenges for local hardware compared to smaller “Flash” models.
Discussions within the community also highlighted the importance of model accessibility for enterprise users. Some commenters noted that compliance requirements, such as FedRAMP, often limit the ability of professionals to use certain international models, underscoring the market demand for specific regulatory-compliant AI solutions. While the technical specifications have been confirmed by the developer, some community members continue to speculate on the potential release of even smaller “Mini” or “Nano” versions of the model.