Moonshot AI Unveils Kimi K3: A Breakthrough in Scale, Performance and Cost for Chinese Frontier Models

Moonshot AI has released Kimi K3, a new frontier model that achieves top-tier performance while significantly reducing inference cost. The model demonstrates competitive scores on benchmarks such as AIME, GPQA, and MMLU-Pro, and uses a sparse Mixture-of-Experts architecture with 44B activated parameters.

Moonshot AI, the Chinese startup behind the Kimi assistant, has unveiled its latest frontier model, Kimi K3. The model is built on a sparse Mixture-of-Experts (MoE) architecture with a total of 123B parameters, of which 44B are activated per token. Kimi K3 achieves state-of-the-art results on several reasoning benchmarks, including 94.1 on AIME 2024, 89.2 on GPQA Diamond, and 79.1 on MMLU-Pro. It also scores 98.6 on MATH-500 and 65.2 on LiveCodeBench. According to Moonshot AI, Kimi K3’s inference cost is only 0.7 yuan per million input tokens, roughly 1 cent, making it significantly cheaper than comparable models. The company claims the model outperforms OpenAI’s o1 and DeepSeek’s R1 on several metrics, while being more cost-effective. Kimi K3 supports a 128K context window and is initially available through an API, with deployment on the Kimi app planned for the future.

Сокращения: MoE = Mixture of Experts — смесь экспертов; AIME = American Invitational Mathematics Examination — Американская пригласительная математическая олимпиада; GPQA = Google-Proof Q&A Benchmark — бенчмарк вопросов и ответов, устойчивый к Google; MMLU = Massive Multitask Language Understanding — массивное многозадачное понимание языка

Source: Moonshot Kimi (GNews) — original