Overview
- Moonshot publicly unveiled Kimi K3 as an open‑weight model that it says approaches the capabilities of leading proprietary systems while acknowledging it still trails the strongest models.
- The company reports Kimi K3 is roughly 2.7–2.8 trillion parameters and supports a one‑million‑token context window, a feature that lets the model process much larger prompts in a single request.
- Moonshot published benchmark results showing Kimi K3 outscoring U.S. rivals on some coding and long‑horizon tests while scoring lower on overall reasoning Elo metrics, but those results come from company reports rather than independent third‑party verification.
- Moonshot advertises substantially lower pricing for input and output tokens than OpenAI and Anthropic, which could cut costs for heavy users but would still require significant hardware to run the model at scale.
- The reveal has intensified debate over provenance and safety: Anthropic has accused some Chinese firms of large‑scale distillation (training on another model’s outputs), U.S. export controls and Chinese policy talks are in play, and the full Kimi weights and independent stress tests have not yet been published for outside inspection.