Moonshot AI has put Kimi back in the center of the open-model race. Kimi K3 launched on July 16 with 2.8 trillion parameters, native visual understanding and a 1M-token context window.
The stronger signal is not just model size. The company has also placed K3 at the top of its official API model list and published usage pricing, turning the release into a service migration rather than a paper launch.
“The larger the parameter count, the higher the capability ceiling, and the smarter the model performance can be,” a Moonshot representative said in the Xinhua report.
The API is already the main entry
Xinhua described Kimi K3 as the largest open-source model by parameter count and said it targets software engineering, knowledge work, deep research and multimodal understanding. Kimi’s own documentation lists kimi-k3 as its most capable model, with 2.8T parameters, vision support and a 1M-token window.
The same official model page says Kimi K2.5 and the Moonshot V1 family are closed to new users after the K3 release, with full platform shutdown scheduled for August 31. That places K3 as the new default route for incoming users.
The price makes the trade-off visible
Kimi’s API homepage lists K3 at $0.30 per million cached input tokens, $3.00 per million regular input tokens and $15.00 per million output tokens. K2.7 Code and K2.6 sit at $0.19 cached input, $0.95 regular input and $4.00 output.
That puts K3 at roughly 3.2x the input price and 3.75x the output price of K2.7 Code. The bet is clear: developers pay more for the 1M-token window, visual input and stronger long-horizon coding or research work.
The open-weight loop still needs evidence
The release has one missing proof point. Official Kimi pages confirm the model name, parameter count, context length, capability direction and API price. The Moonshot organization page on Hugging Face, checked during this run, still showed K2.7 Code, K2.6, K2.5 and K2-series models rather than a K3 weights page.
For builders, the next checks are concrete: weights and license text, active parameter count, a full model card, independent SWE-bench and Terminal-Bench results, and real cost under 1M-token coding or document workloads.
If those pieces land, Kimi K3 moves Chinese open models from cheap alternatives toward large, long-context systems that can compete for serious developer workloads. If they lag, the 2.8T figure remains a strong launch signal but not yet a complete open-model story.
Sources: Xinhua, Kimi API Platform, Kimi official model list, Financial Times, CocoLoop, Moonshot organization page on Hugging Face; checked the Kimi K3 launch node, 2.8T parameters, 1M-token context, vision support, API input/output/cache pricing, K2.5 and Moonshot V1 retirement plan, and public weights-page status.