Thinking Machines released Inkling, its first open-weight model, marking the company’s public debut after spending 18 months building AI infrastructure largely outside public view. (TechCrunch) The release represents the company’s proof point for its thesis that single monolithic models do not serve all use cases optimally.
In efficiency work, PrismML compressed a 27-billion-parameter reasoning model to under 4 GB, small enough to run on an iPhone, with the smallest version retaining 90 percent of original performance and minimal degradation on math and coding benchmarks. (The Decoder) Apple is reportedly already testing the compression technique. In related research, Hugging Face published technical posts on agent building via Shippy and model routing complexity. (Hugging Face Blog - Shippy) (Hugging Face Blog - Model Routing)