Inkling 2 Small

Inkling 2 Small is a multimodal Mixture-of-Experts model developed by Thinking Machines Lab. Released in November 2026, it serves as the first entry in the second generation of the Inkling model family, featuring a large parameter scale designed for advanced reasoning, agentic coding, and complex knowledge workflows. ## Architecture and Technical Details The model features a 48-layer decoder-only transformer backbone integrated with a sparse MoE design containing 594 billion total parameters, of which 17 billion are active per token. Each token is routed dynamically across 8 of 384 available experts. Inkling 2 Small supports native multimodality by combining text processing with a vision encoder that maps pixel-based inputs into a shared hidden space. It handles a context window of up to 1 million tokens and supports variable reasoning effort controls. ## Capabilities and Evaluation Inkling 2 Small demonstrates major performance advancements over its predecessor, particularly in multi-step agentic tasks, coding benchmarks, and hallucination reduction. It employs a selective answering strategy that lowers hallucination rates while supporting extended output generation for complex computational and research-oriented workflows. At launch, the model is accessible via Thinking Machines' developer platform for inference and fine-tuning, with a wider open weights release planned under an Apache 2.0 license.