Inkling-Small: The Lightweight Inkling
Alongside the 975B flagship, Thinking Machines previewed Inkling-Small — 12B active parameters, same training recipe, dramatically lower hardware requirements.
What We Know About Inkling-Small
Thinking Machines previewed Inkling-Small alongside the flagship Inkling AI launch on July 15, 2026. The official description is short but dense: a lighter-weight model with 12B active parameters, trained with a similar recipe to its 975B sibling, achieving strong performance with even lower cost and latency.
That framing matters. "Similar recipe" implies the same 45-trillion-token multimodal pretraining approach and likely the same controllable-thinking-effort design — compressed into a size class that runs on hardware people actually own. Where the full Inkling AI needs at minimum 290GB of RAM+VRAM even at 1-bit quantization, a 12B-active model quantized to 4-bit could plausibly fit in under 16GB.
For most developers watching this launch, Inkling-Small is arguably the more consequential release: the flagship demonstrates the recipe, but Small is the one that will end up on laptops, in fine-tuned vertical products, and inside agents that need cheap, fast multimodal reasoning.
Inkling vs Inkling-Small at a Glance
| Spec | Inkling AI | Inkling-Small |
|---|---|---|
| Active parameters | 41B | 12B |
| Total parameters | 975B | Not disclosed |
| Status | Released (weights on Hugging Face) | Preview announced |
| Training recipe | 45T multimodal tokens | "Similar recipe" (per Thinking Machines) |
| Realistic hardware | 290GB+ RAM/VRAM minimum | Likely single high-end GPU / Apple Silicon |
| Target use | Customization base for orgs | Low-cost, low-latency deployment |
We will update this page as Thinking Machines publishes Inkling-Small's full model card.