Inkling-Small MoE Model Debuts on Text Arena Leaderboard
Inkling-Small, a new Mixture of Experts (MoE) model developed by Thinky Machines, has debuted on the Text Arena leaderboard. It achieved an overall rank of #88 and ranked #21 among open-source models, scoring 1431 points on AutoEval. The model demonstrates high computational efficiency by achieving competitive performance using only 12 billion active parameters out of 276 billion total parameters. This highlights the potential of MoE architectures to deliver high-quality results while minimizing active computing costs. Inkling-Small operates with 12B active parameters out of a total of 276B parameters. Its performance is neck-and-neck with other highly efficient models, such as Minimax M2.7, which scored 1417 with 10B active parameters.
## BACKGROUND
Mixture of Experts (MoE) is an AI architecture that routes inputs to specific specialized sub-models, or "experts," rather than activating the entire neural network for every task. This approach allows models to have a massive total parameter count while keeping the active parameter count—and thus the computational cost per token—relatively low. Text Arena is a benchmark platform that evaluates and ranks LLMs across various text-based tasks like coding, math, and creative writing.