Thinking Machines Lab has released Inkling 2 Small, a sparse Mixture-of-Experts model featuring 594 billion total and 17 billion active parameters. The model accepts both text and image inputs, supports variable reasoning effort, and is currently live on Tinker for inference and fine-tuning. On the Artificial Analysis Intelligence Index, the new model achieves a score of 31, outperforming its predecessor which scored 13.
The architecture comprises a 48-layer decoder-only transformer where each token routes to 8 of 384 experts, alongside a 292-million-parameter vision encoder. Pretraining spanned four phases across massive NVIDIA rack infrastructure, totaling over 12 trillion tokens, followed by a joint reinforcement learning run focused heavily on coding tasks. Thinking Machines Lab plans to publish the weights under an Apache 2.0 license next week following external safety testing, while current access remains restricted to the Tinker platform API.
In benchmark evaluations, Inkling 2 Small scores 86.3% on SWE-Bench Pro v2 Hard and 64.7% on PerceptionBench, though it trails competing models on Terminal-Bench 4.0. The model's improved alignment training successfully reduced internal misbehavior scores from 66.8 down to 14.9. This release matters because it establishes a powerful new multimodal reasoning model for agentic coding and knowledge work, offering developers advanced capabilities with upcoming open-weight availability.