Thinking Machines Lab Launches Inkling Foundation Model
Thinking Machines Lab has released Inkling, a multimodal Mixture-of-Experts foundation model featuring 975 billion total parameters and a 1 million token context window. The model, designed for flexibility and customization via fine-tuning, is available now on the company's Tinker platform. Alongside the main model, the company released Inkling-Small, a lighter-weight 12B active parameter version. Features include controllable thinking effort and a playground interface for developers to interact with the models.
MentionedHugging Face
Research
Research Demonstrates Method for Replicating Expert Financial Judgment in AI
Researchers from Thinking Machines Lab and Bridgewater AIA Labs published findings on a training methodology that enables models to replicate human expert judgment in specialized financial tasks. By utilizing high-quality human-annotated datasets and a training recipe incorporating interleaved batching, CISPO loss, and on-policy distillation, the team created a specialized version of the Qwen3-235B model. This custom model achieved 84.7% accuracy across six financial information filtering tasks, outperforming general frontier models while reducing inference costs by 13.8x. The work demonstrates the efficacy of custom model tuning for specific organizational needs.
