Thinking Machines, the startup founded by ex-OpenAI CTO Mira Murati, has released Inkling-Small, a compressed version of its Inkling language model that maintains near-identical performance at roughly one-quarter the size.
The new model contains 276 billion parameters, a substantial reduction from its predecessor, yet surpasses the original on several benchmarks. This represents a significant efficiency gain in multimodal AI, which processes both text and images. Releasing the model as open source removes licensing restrictions that typically limit commercial deployment.
The rapid iteration cycle, with Inkling-Small arriving just two weeks after the initial Inkling release, reflects Thinking Machines' strategy to build models optimized for practical deployment rather than raw scale. Smaller models reduce computational costs for inference, making them accessible to developers and organizations with limited hardware resources.
The company's focus on efficiency addresses a real problem in AI development. Larger models dominate benchmarks but require expensive GPUs and consume significant power. Inkling-Small's performance parity suggests that architectural improvements and training methodology matter more than parameter count alone.
Murati's involvement signals serious backing for the project. Her previous role at OpenAI positions her to understand both the technical requirements and market demands for production-grade AI systems. The open source release strategy also differentiates Thinking Machines from closed competitors like OpenAI and Anthropic, potentially building developer adoption around a freely available foundation.
The multimodal design matters here too. Models that handle images and text together serve broader use cases than text-only systems, from document analysis to visual reasoning tasks. Performance gains on multiple benchmarks indicate Inkling-Small generalizes well across different problem types.
However, specific benchmark details remain limited from the announcement. The claim that the model "surpasses its predecessor on several benchmarks" needs substantiation through published