Moonshot AI is preparing to release Kimi K3, a model with between 2 trillion and 3 trillion parameters, expected in the coming days as of July 16, 2026. The Beijing-based company's upcoming model is expected to perform at par with or surpass Anthropic's Opus 4.8, according to anonymous sources cited by the Financial Times and reported by TechCrunch.

Kimi K3 will be the largest open-weight AI model from China, and the Financial Times describes it as China's largest AI model to date.

The parameter range places Kimi K3 among the largest models anywhere, open-weight or proprietary. For context, the current generation of frontier models from Western labs has clustered in the hundreds of billions to low trillions of parameters; a 2-to-3-trillion-parameter open-weight release would make the model downloadable and fine-tunable by any organization with sufficient compute to serve it. The open-weight designation means the model's trained parameters will be available for download, as opposed to API-only access. The practical implication is that downstream developers can modify, distill, and deploy the model locally, subject to hardware constraints that remain nontrivial at this scale. Serving a 3-trillion-parameter model in production, even with quantization, requires substantial GPU capacity.

The expectation that Kimi K3 will match or exceed Opus 4.8 sets a specific performance benchmark. Opus 4.8 is Anthropic's current frontier-class model, and closing that gap would place a Chinese open-weight model in direct competition with the top tier of proprietary, API-gated systems. Whether Kimi K3 delivers on this expectation will only be confirmable once independent benchmarks are available after release.

The model's imminent arrival coincides with Moonshot AI's reported efforts to raise fresh capital at a $31.5 billion valuation, per TechCrunch. That figure represents a sharp increase from May 2026, when the company raised $2 billion at a $20 billion valuation. The valuation trajectory, from $20 billion to a reported $31.5 billion target in roughly two months, reflects investor appetite that appears to be tracking the company's model roadmap rather than lagging it. Whether the new round closes at that level remains to be seen; the figure is, at this point, a reported target based on anonymous sourcing.

The Financial Times originally reported the Kimi K3 story on July 16, 2026, citing anonymous sources. Both the Financial Times and TechCrunch, which corroborated the reporting, frame the release as imminent rather than announced through official channels. Moonshot AI has not, per the available reporting, issued a public statement with a specific release date.

Looking at what this means for the open-weight landscape: if Kimi K3 lands at the upper end of the reported parameter range and performs as sources expect, it would narrow the gap between open-weight and proprietary frontier models on both capability and accessibility axes simultaneously. The open-weight community has been operating with capable but smaller models; a 3-trillion-parameter release with competitive benchmark performance would shift the reference point for what "open" can mean at the frontier.

There is also a geopolitical dimension worth flagging. Chinese AI labs have been accelerating open-weight releases, and Kimi K3 would extend that pattern to a new scale. For organizations operating under export-control regimes or with data-residency requirements, an open-weight model of this class, downloadable and self-hosted, offers a path to frontier-tier capability without dependence on U.S.-based API providers. That is a practical consideration for enterprise architects evaluating deployment options, not a speculative one.

The valuation and the model release are, in all likelihood, connected. A model that meets or exceeds Opus 4.8 performance at open-weight scale would provide a concrete justification for a valuation leap from $20 billion to $31.5 billion in a compressed timeframe. Investors pricing a round at that level are betting that Kimi K3's performance claims hold up under independent evaluation.

The coming days will determine whether the reported expectations translate into a released model and, subsequently, whether benchmark results support the performance claims. For now, the verified facts establish intent, parameter scale, open-weight availability, a specific competitive target, and an aggressive valuation context. The rest awaits the release.