Thinking Machines Releases Inkling, an Open-Weights Multimodal MoE Model for Developers

Thinking Machines has launched Inkling, an open-weights multimodal AI model designed for developers building agentic systems, coding assistants, chatbots, and RAG applications. The model accepts text, image, and audio inputs and produces text output, with support for local deployment via libraries such as vLLM, SGLang, and Hugging Face Transformers, as well as API access through third-party providers. Inkling is built on a 66-layer decoder-only transformer with a sparse Mixture-of-Experts architecture featuring 975 billion total parameters, of which 41 billion are active per token. Benchmarks conducted against both open- and closed-weights competitors show strong performance across reasoning, coding, vision, audio, and safety evaluations, including 97.1% on AIME 2026 and 77.6% on SWEBench Verified. Training data was drawn from publicly available sources, third-party acquisitions, and synthetic datasets, with deduplication and safety filtering applied throughout.