Kimi K3 Unleashed: 2.8 Trillion‑Parameter Model Tackles 3D Simulations, Games, and Kaggle

On July 17, Kimi K3 was announced with a staggering 2.8 trillion parameters, making it the largest open‑source large model worldwide. Compared with other models, its size surpasses Moonshot’s 30 trillion‑parameter effort and is far beyond the 1.8 trillion‑parameter DS benchmark.

The first task asked Kimi K3 to produce a Chinese‑style 3D rocket launch‑and‑recovery simulation that runs in a browser. Within ten minutes it returned a complete HTML file with a red‑themed UI, glowing buttons, and a launch‑pad view. The generated simulation displayed staged rocket separation, attitude adjustment, and a soft‑landing sequence, and the author could pause, restart, and drag the camera. After reporting a wobble on the first landing, K3 revised the thrust curve and added camera damping.

The second task was a 3D version of Chrome’s offline dinosaur runner. Kimi K3 generated a low‑poly desert scene with a controllable dinosaur, obstacles, and pterodactyls. After an initial frame‑rate hiccup, the model adjusted input buffering and physics ordering, resulting in smooth gameplay. It also added game‑feel details such as jump‑height curves, landing vibration, scoring sound effects, and a camera zoom on game over—features not explicitly requested.

The third task asked for a web‑based spreadsheet with full formula support. Kimi K3 produced a functional Excel‑like interface with toolbar, cell editing, and formula evaluation. The author verified formulas such as =IF(A1>0, '正数', '非正数'), =IF(AND(A1>0, A2<10), '满足', '不满足'), =VLOOKUP('苹果', A1:C10, 2, FALSE), as well as SUMIF and COUNTIF. Conditional formatting reacted correctly (e.g., cells turning red when values exceeded 50), and the sheet remained stable after copying data and modifying formulas.

For the fourth task, the author instructed Kimi K3 to win the Kaggle “House Prices – Advanced Regression Techniques” competition with a target score of 0.09. Kimi K3 autonomously downloaded the dataset, performed EDA, handled missing values, engineered features, tried several models (Ridge, Random Forest, Gradient Boosting), performed cross‑validation, stacked models, and finally submitted a solution that achieved a public leaderboard score of 0.09 after more than ten hours of computation.

The experiments demonstrate that Kimi K3 can generate complete, runnable front‑end code across diverse domains, but the author notes two major drawbacks: speed (tasks took hours compared to seconds for proprietary models) and aesthetic quality (generated UI designs are functional but lack polished visual appeal). Additionally, the model suffers from knowledge cutoff—failing to answer questions about events after its training data—and the official documentation acknowledges further limitations.

Overall, the author concludes that while earlier Chinese models were viewed as toys or slow prototypes, Kimi K3, along with recent releases from DeepSeek and GLM, shows that open‑source domestic models can now handle realistic engineering tasks, even if they still lag behind top‑tier closed‑source systems in speed and design finesse.