Inkling 2 Small
Thinking Machines Lab · 2026-11-05 · 594B parameters
Inkling 2 Small is Thinking Machines Lab's second-generation multimodal Mixture-of-Experts model, with 594 billion total and 17 billion active parameters. The 48-layer decoder routes each token to 8 of 384 experts, accepts text and image but not audio, supports variable reasoning effort and a context window of up to 1M tokens, and was trained natively in NVFP4 on 12.2 trillion tokens across its first two pretraining phases. It launched on Tinker with closed weights, and Thinking Machines Lab says it intends to release them under Apache 2.0 after external safety testing. Early checkpoints ran on OpenCode as Fledge-Alpha.