Inkling 2 Mini opens its weights: check memory and licensing before local use

The announcement reports roughly 20 GB of model memory use and support for Macs with at least 32 GB RAM. The model card specifies Apache 2.0.

Downloadable weights still need suitable hardware

Thinking Machines Lab released Inkling 2 Mini on November 20, opening its weights and offering inference and customization on Tinker. Its model card specifies Apache 2.0 and text-and-image input with text output. The announcement reports roughly 20 GB of model memory use and initial Ollama support on Macs with at least 32 GB RAM. Model memory use and total system RAM are different measures; neither guarantees smooth operation on every laptop.

Looping does not add stored facts

The 26.4-billion-parameter mixture-of-experts model runs its middle layers three times per token. The company explains that this adds computation rather than factual storage. Researchers demonstrated an offline flight-simulator build aboard an aircraft. That is a vendor demonstration, not our test or evidence that every workflow works offline. Users considering Chinese documents or coding should check their own samples and tool calls before adoption. Tasks that require external search may still use the network even when generation runs locally.

via: Announcement, Model card