NewsBite analysis · AI, Cloud, Data, Programming
MarkTechPost
A 975B-Parameter Open-Weights Multimodal MoE With 41B Active Parameters And Controllable Thinking Effort
Thinking Machines Lab introduced Inkling, their first from-scratch trained model featuring a staggering 975 billion parameters, out of which 41 billion are active. This Mixture-of-Experts (MoE) transformer is designed for customization and supports diverse inputs including text, images, and audio. With robust pre-training on vast datasets, Inkling's architecture optimizes control over reasoning effort and token utilization, making it capable of matching or surpassing previous models on specific evaluations.
ExperimentConfidence: HighShort term horizon
Published 16 Jul 2026, 00:40 UTC4 min read
Visit original article ↗Lead image — MarkTechPost
Summary
TL;DR
- Inkling features 975 billion parameters with 41 billion active.
- Customizable fine-tuning is a core focus of the model.
- It supports a context window of up to 1 million tokens.
- The model is benchmarked against other leading models, maintaining competitive scores.
- Effort control allows for token expenditure adjustments tied to performance.
Interesting facts
- Inkling uses a unique architecture that includes a 66-layer decoder-only transformer.
- The model manages token budgets across rollouts, optimizing efficiency.
- It includes multimodal capabilities, accepting text, audio, and images.