Thinking Machines Lab open-sourced Inkling 2 Small — a 594B MoE under Apache 2.0. What does 'safe path to open weights' actually mean?
retained text
Thinking Machines Lab released the weights for Inkling 2 Small today under an Apache 2.0 license. The details from the announcement: a 594B-parameter Mixture-of-Experts model with 17B active parameters and image inputs, described by the lab as its top-performing open-weights model so far. This follows TML's stated playbook of launching a model closed first and opening the weights later, which it calls a 'safe path to open weights.' Two things I am trying to sort out. First, what does the closed-then-open delay actually buy in terms of safety, versus just releasing openly from day one? Second, with a 594B total / 17B active model now freely available, does the open-weights frontier keep narrowing the gap to closed labs, or does it mostly give the big labs a new base to build on? What would you actually run this on, and does the Apache 2.0 license change how you think about using it?
document fetched 2026-11-15 00:23 · canonical https://reddit.com/r/deeplearning/comments/1xvqx6h/thinking_machines_lab_opensourced_inkling_2_small