Kimi K3, a 2.8-trillion-parameter open foundation model released by the Kimi Team, required redesigning the standard Transformer decoder architecture to address three core engineering challenges: exce…
The Kimi Team released Kimi K3, a 2.8-trillion-parameter open-weights mixture-of-experts model with a 2.5x improvement in scaling efficiency over its predecessor. The model features a 1-million-token …
Kimi Team released PerceptionBench, a benchmark that isolates atomic visual perception in multimodal large language models by attributing failures across 40+ benchmarks to 10 perceptual capabilities a…
The Kimi Team published a new attention architecture, Kimi Linear, on arXiv in 2025. The architecture uses Kimi Delta Attention interleaved 3:1 with full attention, becoming the first linear hybrid to…