Supporting case
Thinking Machines releases Inkling 2 Small’s weights
What happened
On November 12, Thinking Machines updated its November 5 launch to confirm downloadable weights. Its model card ↗ links to the two compressed weight formats and identifies Apache 2.0 licensing. Independent technical reporting found roughly 394 GB for one published download. The model has 594 billion parameters in total and activates about 17 billion per token. Activation describes how much of the model runs for one piece of input; someone hosting it still needs to store the full model and arrange enough memory and processing to serve it. Independent download assessment ↗.
A capable operator can now choose a host other than Tinker. For a small organization the infrastructure cost can still make a hosted service the practical option. This is a real access change from D1107-11’s hosted-only stage, without proof of practical edge deployment.
Dates and implementation stage
2026-11-12: Thinking Machines releases Inkling 2 Small’s weights
