2.8 trillion parameters, $3 per million input tokens, and $15 per million output tokens — in our view, the most important thing about Kimi K3 is not that it is yet another large model launch, but that Chinese vendors are, for the first time, clearly pushing “open large models” into a high-price, high-performance tier.
What this is
Moonshot AI has released Kimi K3, calling it its strongest model to date, and plans to open the weights by July 27, 2026 (that is, make the model parameters public for external deployment and secondary development). Based on the cited data, it performs strongly on long-task knowledge-work benchmarks and frontend code generation, while using fewer output tokens than its predecessor. That suggests this is not just about “scaling up,” but also about improving inference efficiency.
Industry view
The industry will read this as a signal: Chinese large-model companies are no longer competing only on “cheap,” but are instead trying to follow the path of leading overseas players, using stronger capabilities and more complete API pricing to compete for the high end of the market. What matters here is that this approach raises the barrier to enterprise experimentation and shifts competition away from parameter counts toward real task performance.
But the counterarguments are equally valid. First, many benchmark results are vendor-reported and do not necessarily translate into real business outcomes. Second, the reference value of novelty benchmarks such as “a pelican riding a bicycle” is declining, which shows how much harder it is becoming to judge models based on any single test. Third, once prices rise, whether enterprises are willing to keep paying over the long term will depend on stability, delivery, and ecosystem support.
Impact on regular people
For enterprise IT: If K3 follows through on open-sourcing, enterprises will gain another high-performance on-premises option, but budget priorities will shift from “model cost” to deployment, governance, and evaluation.
For individual professionals: Knowledge workers will more quickly see Chinese models that are better at coding and handling long-document organization, but “knowing how to use them” will soon lose to “being able to plug them into workflows.”
For the consumer market: In the short term, users may not immediately feel lower costs; they are more likely to notice stronger products first. The real change is that Chinese AI applications may become more willing to introduce tiered paid offerings.