DeepSeek V4 cuts inference costs with 1.6T parameter model using 13.7x less memory than V3

DeepSeek released V4 on April 24, 2026, offering two open weights models: a 284 billion parameter Flash mixture-of-experts model with 13 billion active parameters, and a 1.6 trillion parameter Pro model with 49 billion active parameters. The company claims the models rival proprietary Western LLMs while dramatically reducing inference costs.

Architecture and efficiency gains

The V4-Pro model was trained on 33 trillion tokens. According to DeepSeek's benchmarks, it outperforms all open weight LLMs and matches leading proprietary models. However, the company's benchmark claims have not been independently verified.

The key technical advancement is a hybrid attention mechanism combining Compressed Sparse Attention and Heavy Compressed Attention. These techniques reduce both compute requirements during inference and memory needed for key-value caches. DeepSeek claims this results in 9.5x-13.7x less memory usage compared to V3.2 while supporting a 1 million token context window.

Both models use mixed FP8 and FP4 precision. The mixture-of-experts weights specifically use FP4 through quantization-aware training—halving memory requirements compared to FP8 at the cost of reduced precision. DeepSeek V3 was among the first open models trained at FP8; V4 pushes further into lower precision territory.

For training, DeepSeek introduced a new optimizer called Muon, designed to speed convergence and improve training stability.

Hardware support and deployment

DeepSeek validated V4 to run on both Nvidia GPUs and Huawei Ascend NPUs. The technical paper confirms the company tested its "fine-grained EP [Expert Parallel] scheme on both Nvidia GPUs and Ascend NPU platforms."

The paper does not specify whether Huawei hardware was used for pre-training or only for inference and post-training reinforcement learning. DeepSeek may have used Nvidia GPUs for initial training and Huawei accelerators for the inference-adjacent reinforcement learning phase.

The models are available on Hugging Face, through DeepSeek's API, and via the company's web service. Pricing has not been disclosed.

What this means

DeepSeek's focus on inference efficiency addresses the primary cost barrier in large language model deployment. Reducing KV cache memory by an order of magnitude while maintaining a 1 million token context window would significantly lower serving costs for providers. The smaller 284B Flash model offers a middle ground between capability and cost.

Validation on Huawei Ascend NPUs is notable given U.S. export restrictions on Nvidia chips to China. If DeepSeek can achieve comparable performance on domestic hardware, it reduces dependency on American semiconductors—though the extent of Huawei chip usage in training versus inference remains unclear. The FP4 quantization strategy suggests DeepSeek is optimizing for hardware with limited precision support or memory constraints.