Author(s): Manpreet Singh

Originally published on Towards AI.

The Least Useful Thing About It.

By the time you read this, your feed will be full of the same headline.

The author argues that although Kimi K3 is genuinely strong and open-weight (ranked #7/190 on Artificial Analysis’s Intelligence Index), “free” is misleading because (1) a 3T model isn’t something most individuals can run locally—it’s effectively free only for companies with GPU clusters, (2) the hosted API is priced above average ($3/M input and $15/M output, with cached inputs cheaper but still not the bargain people assume), and (3) it’s slow and verbose (about 32 output tokens/sec and generates far more tokens than the median), which together can make it dramatically more expensive per real task. The key takeaway is to choose models based on end-to-end cost for the specific workload: K3 is best for long, agentic, unattended runs; workloads needing very large context; situations with heavy cache reuse; and cases where open weights are required for fine-tuning, compliance, or hosting. Conversely, interactive or simple tasks should usually go to cheaper models. The post also advises watching how providers and quantized versions change pricing/speed over the next 48 hours, focusing on agentic benchmarks and verbosity settings, and closes with practical routing guidance and questions about whether most teams actually benefit from using top leaderboard models for everyday tasks.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI