一项针对Kimi K3模型自托管推理的评测显示，该模型在8×B300节点配置下需要1.4TB权重，硬件成本相比采用8×B200节点的GLM-5.2高约20%[1]。尽管投入更高，K3在SWEBench Pro基准测试中的任务完成率达到86.4%，较GLM-5.2和Opus 4.8的62.5%高出24个百分点[1]。在16并发会话场景下，K3的中位任务时间为38分钟，token吞吐量达到122 tok/s[1]。
评测基于对64个SWEBench Pro任务约100次运行的测试数据[1]。研究团队还对多种GPU硬件配置进行了成本效益分析，包括DGX Spark、H200、HGX H200及HGX B200等[1]。分析表明，即使B200 GPU机架仅保持15%的利用率，其成本效益也能超越前沿API定价模式[1]。
根据调查数据，中位员工年度AI API使用支出约为140美元，第90百分位数接近7,300美元，第99百分位数接近90,000美元[1]。主要编码用例占大型模型提供商年度经常性收入的70%以上[1]。
A performance evaluation of the Kimi K3 model demonstrates that self-hosting the system requires approximately 20% more hardware investment than its predecessor, yet delivers substantially better results on coding tasks [1]. The K3 model requires 1.4TB of weights across 8×B300 GPU nodes, compared to the 8×B200 configuration needed for GLM-5.2, increasing infrastructure costs accordingly [1].
Despite the higher hardware requirements, K3's capabilities justify the investment for certain enterprise applications [1]. On the SWEBench Pro benchmark, K3 achieved a task completion rate of 86.4%—a 24 percentage point improvement over both GLM-5.2 and Opus 4.8, which each scored 62.5% [1]. The model processed 16 concurrent sessions with a median task time of 38 minutes and maintained a token throughput of 122 tokens per second [1]. These metrics were derived from approximately 100 test runs across 64 SWEBench Pro tasks [1].
The economics of self-hosting versus cloud API consumption increasingly favor on-premises infrastructure for high-volume users [1]. A B200 GPU rack operating at just 15% utilization can outprice leading API services [1]. Employee-level AI API expenditure analysis reveals a median annual cost of $140 per worker, with the 90th percentile reaching approximately $7,300 and the 99th percentile approaching $90,000 [1]. Given that coding applications account for over 70% of annual recurring revenue for major large language model providers, organizations with substantial development teams face a potential inflection point for shifting from pay-per-use cloud models to self-hosted inference infrastructure [1].