DeepSeek V4

DeepSeek

Open-weight model · runs on-premises · Published on 5 July 2026 · Updated on 2 September 2026

DeepSeek V4 delivers frontier-level performance at reduced cost. The Pro version totals 1.6 trillion parameters (49 billion active), the Flash version 284 billion (13 billion active), both under the MIT licence. The Pro version handles reasoning, while the Flash version serves everyday tasks with speed and fits on a single-node machine. This model forms the foundation of dwarfstar.

Key points: Price-to-performance ratio · Mixture of experts · Pro 1.6T, 49B active · Flash 284B, 13B active · Foundation of dwarfstar.

Architecture

DeepSeek V4 is a family of two mixture-of-experts models, both with a native context of 1,048,576 tokens and an MIT licence. DeepSeek V4 Pro totals 1,600 billion parameters of which 49 billion are active per token (384 experts, 6 active), for 865 GB of weights published in FP4 and FP8. DeepSeek V4 Flash totals 284 billion parameters of which 13 billion are active (256 experts, 6 active), for 167 GB of weights published in the 0731 revision. Both figures come from the official model card on Hugging Face, re-read on 2 September 2026.

Strengths

It delivers frontier-level performance at reduced cost, making it a benchmark for price-to-performance among open-weight models.

Use cases

The Pro version handles demanding reasoning tasks, while the Flash version serves everyday work with speed. DeepSeek V4 forms the foundation of the dwarfstar stack.

Deployment

Served by vLLM or llama.cpp depending on quantisation. It runs locally on an RTX PRO 6000 server within dwarfstar.

All models on the platform are interchangeable through a single gateway: switching models is a matter of changing one configuration line, with no code rewrite. Local execution remains the priority.

Official sources

• deepseek-ai/DeepSeek-V4-Flash-0731, official model card on Hugging Face: licence, parameters, context window and weight format
• deepseek-ai/DeepSeek-V4-Pro, official model card on Hugging Face: 1.6T parameters, 49B active
• Artificial Analysis, model comparison index