DeepSeek releases 284‑billion‑parameter V4 model for local laptop use

DeepSeek unveiled its V4 Flash model, a 284‑billion‑parameter large language model that can run on high‑end laptops. Using a Mixture‑of‑Experts architecture, advanced quantization and the DwarfStar DS4 inference engine, only about 13 billion parameters are activated per token, allowing execution on devices such as the MacBook Pro M3 Max or Nvidia DGX Spark with a 76 GB memory footprint. The model is released under an MIT licence, promoting open‑weight access while requiring premium hardware and acknowledging trade‑offs in accuracy and accessibility.

Separately, DeepSeek published a developer guide for its V3 API, which offers OpenAI‑compatible endpoints, a 64 K‑token context window, and competitive performance on code‑generation and reasoning benchmarks at a lower per‑token cost than leading rivals. The guide details setup, streaming, error handling and migration from other providers, highlighting the model’s Mixture‑of‑Experts design and cost‑effective pricing.

Both releases emphasize democratizing AI by providing powerful, open‑weight models that can be deployed locally or accessed via affordable APIs.