DownloadInkling-Small is an open-weight general-purpose multimodal model from Thinking Machines Lab, designed for agentic systems, coding assistants, chatbots, retrieval workflows, and natural-language applications. It accepts text, images, and audio as input and produces text output, with multilingual and multi-programming-language capabilities. The model uses a sparse Mixture-of-Experts architecture with 276B total parameters and 12B active per token, enabling strong performance with lower inference cost than a fully dense model of similar scale. Its 42-layer decoder routes each token through six of 256 specialized experts plus two shared experts, while hybrid local and global attention supports efficient processing. Inkling-Small performs strongly across software engineering, tool use, mathematics, vision, and audio benchmarks, including 80.2% on SWE-Bench Verified and 95.5% on AIME 2026.
Features
- 276B total parameters with 12B active per token
- Native support for text, image, and audio inputs
- Sparse Mixture-of-Experts architecture with 256 experts
- Hybrid local and global attention layers
- Designed for agents, coding assistants, chatbots, and RAG
- Strong coding, mathematical, vision, and audio performance
- Supports BF16 and memory-efficient NVFP4 deployment
- Compatible with Transformers, vLLM, SGLang, and Unsloth
Project Samples
