Kimi K3.1 Pro Tops Open-Weight Models Amid Paid Marketing Allegations

Chinese startup Moonshot has launched the Kimi K3.1 Pro model, a post-trained version of the Kimi K3 model released in July, running on three trillion parameters. The company stated that it will release the model weights under an open license on November 4, which could make it the world's most powerful open-weight model based on benchmark results.

The model scored 55 points on the Intelligence Index by independent testing firm Artificial Analysis, ranking fourth overall, ahead of models like Xiaomi MiMo-V2.6-Pro, DeepSeek V4.1-Flash-1017, Meta Muse Wonder, and even OpenAI's GPT-6 Astra. The largest jump came in the Terminal-Bench 4.0 benchmark, rising from 13% to 67%—a 54 percentage point increase—outperforming Sonnet 5.5, which scored 64% in the same firm's evaluation.

The model excels at multi-step sequential tasks such as maintaining software repositories, debugging, and test writing, in addition to financial analysis, 3D modeling generation, and software security work. Developers can access it via applications and an API at prices nearly half those of Claude Opus 5.5 and Meta Muse Wonder, though DeepSeek and Xiaomi remain cheaper.

Conversely, the launch was accompanied by controversy on X, where allegations were raised regarding an undisclosed paid promotional campaign by influencers. Moonshot did not immediately issue a clarification regarding these accusations. Benchmark results also show limited gains in certain areas; for instance, Humanity’s Last Exam only rose from 47% to 52%, suggesting that the largest improvement was in task execution rather than general reasoning ability.

What do these terms mean?

Parameter: A numerical value within the model's network that is adjusted during training; the higher the number, the greater the model's ability to represent complex patterns.

Open-weight: A model published alongside its weight files so others can run or modify it, though this does not necessarily mean the training data is published.

Intelligence Index: An aggregate metric issued by independent entities to compare model performance across multiple benchmarks, which should be read alongside the details of each individual test.