NewsBite analysis · AI, Cloud, Startups
MarkTechPost
A 276B Total, 12B Active Open Weights Multimodal MoE Model
Thinking Machines Lab has unveiled Inkling-Small, an open weights Mixture-of-Experts model designed to run efficiently on NVIDIA systems. With 276B total parameters and 12B active weights, it reduces hardware requirements for businesses looking to deploy AI in regulated sectors. The model supports audio, image, and text inputs, enabling diverse applications, especially in analytics and automation. Importantly, Inkling-Small surpasses its predecessor Inkling in key performance benchmarks, offering enhanced reasoning capabilities.
MonitorConfidence: HighShort term horizon
Published 2 Aug 2026, 20:40 UTC4 min read
Visit original article ↗Lead image — MarkTechPost
Summary
TL;DR
- Inkling-Small features 276B total parameters and 12B active weights.
- The model runs efficiently on NVIDIA B300 and H200 hardware.
- Applicable to regulated sectors like finance, healthcare, and telecom.
- Trained on diverse inputs, including text, images, and audio.
- Outperformed its larger predecessor in several key benchmarks.
Interesting facts
- Each token routes through 6 of 256 experts in the model architecture.
- It can run on a single rented B300 instance, making it accessible for startups.
- The model combines various modalities, enhancing its applicability beyond text-only tasks.