What changed
Thinking Machines Lab released Inkling 2 Small on Tinker for playground use, sampling and fine-tuning. The 594B-total, 17B-active multimodal MoE accepts text and images, offers variable effort and up to a 1-million-token context; downloadable Apache 2.0 weights were promised only after external safety testing.
Why it matters
The hosted release creates a concrete model and fine-tuning path for coding, vision and long-context work while preserving a clear line between today's closed-weight access and a planned open artifact.
What to do
Run a held-out coding and visual task in Tinker, pin effort and context settings, measure quality, latency, token use and fine-tuning cost, and wait for the exact weight checkpoint, license and external safety results before planning self-hosting.
Caveat
The launch benchmarks and safety results are provider-reported, several comparisons use private or versioned evaluations, and the announced Apache 2.0 weights were not yet available on 5 November. Hosted access is governed by Tinker terms.