DeepSeek released the preview of its fourth-generation large language model, DeepSeek V4, on April 24, 2026. The model family includes two variants: V4-Pro and V4-Flash. V4-Pro contains 1.6 trillion total parameters and activates 49 billion per token. V4-Flash contains 284 billion total parameters and activates 13 billion per token. These models compete with closed-source models like GPT-5.4, Claude Opus 4.6, and GPT-5.5. DeepSeek V4 provides superior reasoning performance for a fraction of the cost of closed-source models.
New architectures for large-scale deployment
The DeepSeek architecture uses several new components to handle large-scale training and inference. DeepSeek incorporates Manifold-Constrained Hyper-Connections (mHC) to maintain stability. The system also uses a hybrid system combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). The Engram Conditional Memory module provides O(1) knowledge retrieval. This module separates factual memory from computational reasoning. This separation occurs for the first time in a production model. DeepSeek V4 also uses a Mixture-of-Experts (MoE) architecture. This design activates only a small subset of the model’s total parameters for each input token. This approach decouples model capacity from inference cost. The model stores knowledge across its full parameter set, but only engages relevant experts for computation. This keeps per-token costs low.
Scaling stability through mHC
DeepSeek researchers published papers on mHC between December 2025 and January 2026. mHC addresses signal propagation instability in deep neural networks. Standard residual connections use scalar addition, which causes issues as models grow to hundreds of layers. mHC replaces these connections with learned linear combinations across multiple input streams. The implementation projects learned residual mixing matrices onto the Birkhoff Polytope. This mathematical space of doubly stochastic matrices ensures every row and column sums to 1. This constraint ensures lossless signal propagation. Information flows through the network without being amplified or destroyed. This guarantees the identity mapping is preserved and prevents both gradient vanishing and gradient exploding in ultra-deep networks. The implementation uses the Sinkhorn-Knopp algorithm from optimal transport theory to enforce the doubly stochastic constraint with approximately 20 iterations per forward pass. DeepSeek V4-Pro uses only 49 billion active parameters for any given token, which means the actual computational cost per inference step remains close to that of a dense 49 billion parameter model despite its massive 1.6 trillion parameter scale.
Efficiency in attention and context
The hybrid attention system interleaves CSA and HCA across model layers. This system allows V4 to process 1-million-token contexts efficiently. V4-Pro reduces overall inference FLOPs by 73% compared to V3.2. V4-Flash achieves approximately a 90% reduction in inference FLOPs. For a 1-million-token context, V4-Pro uses only 10% of the KV cache used by V3.2. V4-Flash uses only 7% of the V3.2 cache. The Engram module also improves long-context performance. The Engram technology enables efficient retrieval from contexts exceeding one million tokens.
Hardware and the Chinese ecosystem
DeepSeek V4 runs on Huawei Ascend 950PR chips. This makes it the first frontier-class AI model built entirely on Chinese domestic semiconductor infrastructure. This capability shows that world-class AI models no longer require NVIDIA hardware. DeepSeek, formally Shenzhen DeepSeek AI Technology Co., Ltd., was founded in 2023 by Liang Wenfeng. Liang Wenfeng was a hedge fund manager before becoming an AI entrepreneur. The company targets a mission to build world-class models and release them under permissive licenses. This approach drives down the cost of inference to levels that make advanced AI accessible to every developer, startup, and researcher.
Reasoning modes and benchmark performance
The models support three reasoning effort modes: Non-think, High, and Max. Non-think mode bypasses extended chain-of-thought reasoning for fast, intuitive responses. High mode activates standard reasoning chains for complex problem-solving. Max mode allocates the maximum compute budget for deep reasoning. In Max mode, V4-Pro scores 91.2 on MMLU-Pro and 90.1 on GPQA Diamond. It also achieves a Codeforces Elo rating of 3,206, a score that places the model ahead of 96.3% of human competitive programmers. V4-Pro remains slow during complex multi-turn tasks with heavy tool use.
| Model | Total Parameters | Active Parameters | MMLU-Pro (Max) | Codeforces Elo |
|---|---|---|---|---|
| DeepSeek V4-Pro | 1.6T | 49B | 91.2 | 3,206 |
| DeepSeek V4-Flash | 284B | 13B | 86.2 | 2,816 |
The 1-million-token context window is functionally usable. V4-Pro achieves 97% accuracy on the Needle-in-a-Haystack test at full 1M token context length. This makes it practical for full-codebase analysis and long-form agent workflows. V4-Pro manages project-wide logic across large codebases due to multi-file reasoning. It also maintains coherence over significantly longer prompts.
Economics of the V4-Flash variant
V4-Flash provides frontier-tier performance at a lower cost. V4-Flash achieves an Artificial Intelligence Index score of 47 in Max mode. This puts it at the same level as Claude Sonnet 4.6 (Max mode). V4-Flash costs $0.28 per million output tokens. V4-Flash input costs $0.14 per million cache-miss tokens and $0.028 per million cache-hit tokens. V4-Flash won 7 out of 20 real-world tasks at a cost of $0.04. You should consider the efficiency gains if your workload involves massive datasets. Will the Engram architecture allow for infinite context retrieval?
DeepSeek’s trajectory and industry impact
DeepSeek is preparing to raise $7.4 billion in its first external funding round. This fundraising could value the company between 350 billion yuan and 400 billion yuan. Liang Wenfeng expects to contribute 20 billion yuan of his own capital. Tencent Holding Ltd. considers an investment of about 10 billion yuan. CATL evaluates a 5 billion yuan commitment. Other prospective investors include NetEase, Inc., JD.com, IDG Capital, and Monolith Capital. DeepSeek released V2 in May 2024. V2 introduced Multi-Head Latent Attention (MLA) and reduced the KV cache by 93.3%. DeepSeek V3 followed in December 2024. V3 scaled to 671 billion total parameters. DeepSeek trained V3 for approximately $5.6 million. DeepSeek R1 arrived in January 2025. It showed that reinforcement learning could produce powerful reasoning capabilities.
