DeepSeek has officially launched its V4 model family, marking a significant push in the AI efficiency race. The lineup includes two variants: V4-Pro with 1.6 trillion total parameters (49 billion active per token) and V4-Flash with 284 billion total parameters (13 billion active). Both are text-only and support a 1 million token context window, positioning them as direct competitors to models like OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.8.
The architecture introduces two key innovations: Manifold-Constrained Hyper-Connections (mHC), which stabilizes training of massive MoE models by limiting signal amplification to under 2x (versus an unconstrained 3000x), and DeepSeek Sparse Attention (DSA), which compresses context into coarse summaries and applies full attention only to relevant regions, enabling near-linear cost scaling for long contexts.
Pricing is a major differentiator. V4-Pro costs $0.435 per million input tokens (cache miss) and $0.87 per million output tokens, while V4-Flash is $0.14 and $0.28 respectively. This is 5.7x to 28.7x cheaper than GPT-5.4 and Claude Opus 4.8 on input and output. The models are open-sourced under the MIT license, and DeepSeek claims they can run locally on dual RTX 4090s or a single RTX 5090, though the full 1.6T model remains a data-center workload.
DeepSeek has also optimized V4 for domestic Chinese silicon, including Huawei Ascend 950PR and Cambricon MLU chips, due to US export restrictions on advanced Nvidia GPUs. The Ascend 950PR reportedly delivers 2.87x the compute performance of the Nvidia H20. This marks a strategic shift toward AI semiconductor independence in China.
Benchmark figures cited in the source are based on leaked internal data and await third-party verification. The company also retired legacy model IDs 'deepseek-chat' and 'deepseek-reasoner' on July 24, 2026.