FULL STORY

Inkling 2 Small Goes Open-Weight

Thinking Machines open-sources the weights of Inkling 2 Small, and vLLM quickly ships a one-command deployment recipe for the 594B MoE model with million-token context.

2026-11-13 ~ 2026-11-13 · 2 episodes · 13 posts

Episode 1 · Thinking Machines Open-Sources Inkling 2 Small Weights Under Apache 2.0 (2026-11-13, 11 posts)

Thinking Machines founder Miriam Murati announced the release of the Inkling 2 Small model weights as promised. The model is now available on Hugging Face under an Apache 2.0 license, allowing the community to fine-tune and use it commercially. This fulfills the team's commitment to open weights for its small model—a notable example of a top startup releasing a large-parameter model under a permissive license.

Confirmed

  • The model is Inkling 2 Small, with weights available on Hugging Face under Apache 2.0 (announced by @miramurati and @thinkymachines)
  • Architecture is MoE, with 594B total parameters and 17B active parameters (@ariG23498 citing the official HF article)
  • Supports adjustable reasoning via a custom chat template
  • HF updated its official model article accordingly; HF engineers including ariG23498 reshared the announcement
  • In her announcement, Murati thanked the team and safety-testing partners; members such as Andrew Carr congratulated the release

Why it matters

  • Apache 2.0 is among the most permissive open-source licenses, allowing commercial use and free fine-tuning, which helps grow the community ecosystem
  • Releasing a 594B-parameter MoE with open weights makes it one of the larger-scale open models, with real reference and practical value for the open-source community

Episode 2 · vLLM adds Thinking Machines' Inkling-2-Small with 1M context and 1100 tok/s (2026-11-13, 2 posts)

vLLM released a one-command deployment recipe for Thinking Machines' Inkling-2-Small, a 594B-parameter MoE with 17B active, 1M context, and 1100 tok/s on a single GPU, along with DSpark speculative decoding details.