Inkling-Small Debuts at Rank #88 Overall on LMSYS Text Arena
Thinky Machines' open-weights model, Inkling-Small, has debuted at rank ~#88 overall and ~#21 among open-source models on the LMSYS Text Arena. This initial ranking is based on early AutoEval scores, where the model achieved 1,431 points. Inkling-Small is notable as one of only five US-developed models to rank in the top 25 open models on the leaderboard. Its strong performance highlights the competitiveness of US-based open-source AI development against dominant international models. Inkling-Small is a 276-billion parameter Mixture of Experts (MoE) model with 12 billion active parameters. The ranking is based on LMSYS's AutoEval, an automated evaluation framework that uses a reward model trained on human preference data rather than direct human votes.
## BACKGROUND
LMSYS Chatbot Arena is a widely respected benchmarking platform where LLMs are evaluated through crowdsourced, blind pairwise comparisons. AutoEval is an automated evaluation framework designed to approximate human preferences using LLM judges and neural metrics. Thinky Machines recently released the Inkling family of open-weights models, which utilize Mixture of Experts (MoE) architectures to optimize computational efficiency.