Bijan Bowen reviews Inkling, a roughly 975-billion-parameter mixture-of-experts model from Thinking Machines with 41 billion active parameters. The Apache-licensed model accepts text, images and audio, emphasizes customization and is available through hosted access and downloadable quantizations.
Bijan Bowen tests browser interfaces, 3D scenes, a C++ game, front-end design, image-guided coding and creative roleplay. Inkling often needs corrections, produces limited visual detail and falls short of the stronger coding models used as comparison points, although it demonstrates some multimodal troubleshooting and creative range.
Bijan Bowen concludes that Inkling's best fit may be enterprise deployment rather than hobbyist coding. A US-based organization could fine-tune it on private data and reduce reliance on expensive closed-model APIs, while a planned smaller version may eventually be more practical for local hardware.
Watch the original on YouTube