Kimi K3 编程测试 (2026)
Tested Kimi K3 for Coding

原始链接: https://www.dotnetperls.com/2026_7_18_tested-kimi-k3-coding

新发布的 Kimi K3 模型被誉为对标 Fable 和 GPT 的前沿级产品,近期对其优化复杂 Rust 函数的能力进行了测试。在被要求改进代码时,该模型建议使用 SWAR(寄存器内 SIMD)技术并行处理 8 位数值。 Kimi K3 在约 15 分钟内成功完成了优化,且未出现任何错误。最终代码使程序性能提升了 2%。考虑到它仅以 1 美元的 token 成本就准确识别并执行了复杂的优化,作者认为 Kimi K3 的性能确实与其它顶级 AI 模型相当。

抱歉。
相关文章

原文

The new model Kimi K3 was released this week, and it is claimed to be near-frontier level quality similar to Fable and GPT. I decided to try it out and see if it could optimize my Rust program. I felt a less specific task, like "optimize this function somehow" was a good test.

Kimi K3

I set up Kimi K3 in OpenCode, and made a copy of my program directory so that I could throw away the changes made easily. I directed a prompt to the function I wanted optimized. I had spent a fair amount of time optimizing this function, and I did not seem to be making much progress on my own.

Kimi K3 had 2 recommendations, and I decided to have it pursue the SWAR optimization—SIMD Within a Register. With SWAR, a single 64-bit value is treated as 8 eight-bit values—so we can handle 8 u8 values (single-byte) at once.

Kimi K3 took about 15 minutes for the task, and it completed it perfectly. I found the optimization sped up the program by 2%. So Kimi K3 both decided on an appropriate optimization, and implemented it with no errors—it seems to me that Kimi K3 could well be a frontier-level model as advertised. It cost about $1 for the tokens on OpenRouter.

联系我们 contact @ memedata.com