DeepSeek's new models offer big inference cost savings

"DeepSeek V4 is available in two flavors: a smaller 284 billion parameter Flash mixture-of-experts model and a larger 1.6 trillion parameter model, with 49 billion active parameters at any moment."

"V4-Pro was trained on 33 trillion tokens and claims to outperform every open weight LLM while rivaling the best proprietary models across its benchmark suite."

"DeepSeek V4 introduces several novel architectural changes that should make the model much less expensive to serve, including a second smaller Flash model for a more interactive user experience."

DeepSeek V4 features two models: a 284 billion parameter Flash mixture-of-experts model and a 1.6 trillion parameter model. The V4-Pro model was trained on 33 trillion tokens and claims to outperform open weight LLMs while competing with proprietary models. Despite its promising specifications, skepticism remains regarding real-world performance. The new architecture aims to lower operational costs, and the introduction of a smaller model enhances user experience and affordability. DeepSeek's previous models have established a strong reputation, but benchmarks may not fully reflect practical applications.

Read at Theregister