Is DeepSeek v4 (Flash) really extremely cheap to run? If yes, how?
Qwen 相关模型动态已经出现，适合跟踪能力变化、生态影响和后续可用性。
DeepSeek v4 (Flash) 模型因其极低的运行成本引发了讨论。尽管该模型拥有 284B 参数，但其提供商（例如 OpenRouter）的收费却异常低廉，甚至比参数量仅为其十分之一的 27B Qwen 模型还要便宜。这让用户感到困惑，质疑提供商是否在进行倾销，或者 DeepSeek v4 的架构是否在服务成本方面存在独特之处。…
Hi. I don't have a GPU. So my biggest "local LLM" experience has been running ~26B models with single-digits tps values.
However, the "serving economy" of DSv4 models look like a riddle to me. The Flash model has 284B parameters, but providers (e.g. OpenRouter) charge so little for it it's ridiculous. It's for example cheaper than 27B Qwen, A tenth of its (total) size! How is it viable?
Are the providers just doing dumping here? Or is DSv4 architecture somehow different in making it extremely cheaper to serve? Those of you who have had the equipment to host DSv4, is there something making it different?