← November 13, 2026 briefing · added in the 00:00 KST update

Inkling 2 Small follow-up: officially added to Hugging Face Transformers v5.23.0, with 48 MoE layers routing to 8 of 384 experts each

Announced November 13, 2026

Verified (fact)

Hugging Face released Transformers v5.23.0, adding Thinking Machines Lab's 'Inkling 2 Small' to its newly supported models. According to the release notes, the model takes text and images as input and outputs text, and is designed to handle multiple languages beyond English as well as multiple programming languages. The language backbone consists of 48 MoE decoder layers, each routing tokens to 8 of 384 experts. It has 594B total parameters and 17B active parameters. It keeps the original Inkling decoder's learned relative attention, short causal convolutions, and MoE layer setup, and adds block-level attention residuals and gated attention with learned sinks. Whether the weights have been released cannot be confirmed from this material.

Why it matters (interpretation)

Official support in the widely used Transformers library lays the groundwork for companies to build custom applications right away once the previously announced weights are released. The official documentation also discloses architecture details for the first time, making it possible to compare its structure with other large MoE models.

Sources

Story thread

  1. 2026-11-06 Thinking Machines Lab releases 'Inkling 2 Small', a 594B-total, 17B-active MoE with image input and long autonomous tasks (community reports)
  2. 2026-11-07 Inkling 2 Small follow-up: Thinking Machines Lab plans to release weights for enterprise customization (report)
  3. 2026-11-13 Inkling 2 Small follow-up: officially added to Hugging Face Transformers v5.23.0, with 48 MoE layers routing to 8 of 384 experts each

More about Thinking Machines Lab