On April 24, 2026 (21 hours ago), DeepSeek released the preview of its fourth-generation large language model: DeepSeek V4. Available in two distinct variants — V4-Pro and V4-Flash — the release represents a watershed moment in the open-source AI landscape. With 1.6 trillion total parameters (49 billion active per token) in its flagship model, DeepSeek V4-Pro achieves performance that rivals the world’s top closed-source models, including GPT-5.4, Claude Opus 4.6, and GPT-5.5, at a fraction of the cost.
The architectural innovations are equally significant. The model introduces Manifold-Constrained Hyper-Connections (mHC) for training stability at extreme scale, a hybrid Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) system that reduces inference FLOPs by up to 73%, and the Engram Conditional Memory module that provides O(1) knowledge retrieval, separating factual memory from computational reasoning for the first time in a production model.
Perhaps most notably, DeepSeek V4 runs on Huawei Ascend 950PR chips, making it the first frontier-class AI model built entirely on Chinese domestic semiconductor infrastructure. This report provides a comprehensive analysis of DeepSeek V4, covering its architecture, benchmark performance, cost dynamics, real-world applications, and broader industry implications.
DeepSeek, formally known as Shenzhen DeepSeek AI Technology Co., Ltd., was founded in 2023 by Liang Wenfeng, a former hedge fund manager turned AI entrepreneur. From its inception, the company pursued a mission that diverged sharply from the prevailing industry consensus: build world-class AI models and release them as open source under permissive licenses, while driving down the cost of inference to levels that would make advanced AI accessible to every developer, startup, and researcher on the planet.
The company’s trajectory has been remarkable. DeepSeek V2, released in May 2024, introduced the Multi-Head Latent Attention (MLA) mechanism and a fine-grained Mixture-of-Experts (MoE) architecture that reduced the KV cache by 93.3%. DeepSeek V3 followed in December 2024, scaling to 671 billion total parameters with 37 billion active per token, trained for approximately $5.6 million — a cost that sent shockwaves through an industry accustomed to nine-figure training budgets. In January 2025, DeepSeek R1 demonstrated that pure reinforcement learning could produce powerful reasoning capabilities without extensive supervised fine-tuning.
DeepSeek V4 is not an incremental update; it is a generational leap. The model family ships in two variants, V4-Pro and V4-Flash, both built on a Mixture-of-Experts architecture but targeting dramatically different use cases and price points. V4-Pro, with its 1.6 trillion total parameters, is the flagship, designed to compete directly with the most powerful closed-source models in the world. V4-Flash, at 284 billion total parameters (13 billion active), is optimized for speed and cost, delivering frontier-tier performance at a price point that makes it viable for the most demanding high-throughput workloads.
The release carries immense geopolitical significance. By running natively on Huawei Ascend 950PR chips, DeepSeek V4 demonstrates that world-class AI models no longer require NVIDIA hardware. This is the first frontier-class model trained and deployed entirely on Chinese domestic semiconductor infrastructure, a milestone that has profound implications for the global AI supply chain, U.S. export controls on advanced chips, and the broader competition between American and Chinese AI ecosystems. Reuters, CNBC, Fortune, and MIT Technology Review all covered the release, with MIT Technology Review highlighting DeepSeek’s “significantly enhanced agentic capabilities” as one of three reasons why V4 matters.
DeepSeek V4 is built on a Mixture-of-Experts (MoE) architecture, a design paradigm that 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 can store an enormous amount of knowledge across its full parameter set, but only the relevant experts are engaged for any given computation, keeping per-token costs low.
V4-Pro is DeepSeek’s most capable model ever built. With 1.6 trillion total parameters, it represents one of the largest neural networks ever deployed in production. However, the MoE architecture ensures that only 49 billion of those parameters are active for any given token, meaning the actual computational cost per inference step is closer to that of a dense 49B parameter model. This is the magic of MoE: V4-Pro has the knowledge capacity of a 1.6-trillion-parameter model but the inference cost profile of a model roughly thirty times smaller.
The model supports three reasoning effort modes:
Non-thinkmode bypasses extended chain-of-thought reasoning for straightforward tasks where speed matters most.Highmode (the default) activates standard reasoning chains, suitable for most applications.Maxmode allocates the maximum compute budget for deep reasoning, achieving the highest benchmark scores but at greater latency and cost.
In Max mode, V4-Pro scores 91.2 on MMLU-Pro, 90.1 on GPQA Diamond, 93.5 on LiveCodeBench, and achieves a Codeforces Elo rating of 3,206, outperforming 96.3% of human competitive programmers.
V4-Pro’s 1-million-token context window is functionally usable. The model achieves 97% accuracy on the Needle-in-a-Haystack test at full 1M token context length, meaning it can reliably retrieve a specific piece of information embedded anywhere within a million tokens of text. This makes it practical for full-codebase analysis, comprehensive document review, and long-form agent workflows.
If V4-Pro is DeepSeek’s answer to the world’s most powerful models, V4-Flash is something far more disruptive: a frontier-tier model priced so low that it changes the economics of AI deployment entirely. At 284 billion total parameters with 13 billion active per token, V4-Flash is substantially smaller than V4-Pro, but its performance is surprisingly competitive.
In Max reasoning mode, V4-Flash achieves an Artificial Analysis Intelligence Index score of 47, placing it at the same level as Claude Sonnet 4.6 (Max mode) — a model that costs nearly 90 times more per output token. DeepSeek confirms that “V4-Flash-Max achieves comparable reasoning performance to the Pro version when given a larger thinking budget, though its performance gaps slightly on more complex tasks.”
In practical testing, V4-Flash in its default mode won 7 out of 20 real-world tasks outright, including five coding tasks, at a total cost of just $0.04 for the entire evaluation. This makes it the preferred choice for high-volume, cost-sensitive workloads where near-frontier performance is sufficient and per-query cost is the primary concern.
V4-Flash also excels at simple agent tasks. DeepSeek states that it “performs on par with V4-Pro on simple Agent tasks,” making it an ideal backend for tool-calling pipelines, multi-turn conversational agents, and automated workflows. At $0.14 per million cache-miss input tokens (and as low as $0.028 per million cache-hit tokens), it is the cheapest frontier-tier model available on the market, full stop.
DeepSeek V4 is not merely a scaled-up version of V3. It introduces at least four major architectural innovations, each supported by dedicated research papers published between December 2025 and January 2026. These innovations address fundamental challenges in scaling large language models: training stability at extreme depth, efficient long-context processing, separation of factual memory from reasoning computation, and advanced reinforcement learning for reasoning enhancement.
The first major innovation is Manifold-Constrained Hyper-Connections (mHC), detailed in a paper published in December 2025 (arXiv:2512.24880) that has since been cited by over 33 papers. mHC addresses a fundamental problem in deep neural network training: as models grow to hundreds or thousands of layers, standard residual connections (simple addition of layer input to layer output) create signal propagation instability.
Hyper-Connections generalize residual connections by replacing scalar addition with learned linear combinations across multiple input streams, but unconstrained Hyper-Connections suffer from runaway signal amplification, causing training to diverge at extreme scale.
mHC solves this by projecting the learned residual mixing matrices onto the Birkhoff Polytope — the mathematical space of doubly stochastic matrices. A doubly stochastic matrix has two critical properties: every row sums to 1, and every column sums to 1. This constraint ensures lossless signal propagation: information flowing through the network is redistributed across layers but is never amplified or destroyed, guaranteeing that the identity mapping is preserved and preventing both gradient vanishing and gradient exploding in ultra-deep networks.
The implementation uses the Sinkhorn-Knopp algorithm, a technique from optimal transport theory, to enforce the doubly stochastic constraint with approximately 20 iterations per forward pass. IBM, Fireworks AI, and multiple research groups have flagged mHC as a foundational contribution to scaling stability, with potential applications extending beyond attention to MoE gating, layer-wise learning rates, and attention mixing. In V4, mHC replaces standard residual connections entirely across the entire model.
The second major innovation is V4’s hybrid attention system, which interleaves two novel attention types across the model layers: Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). This system is the primary engine behind V4’s ability to process 1-million-token contexts efficiently, and it delivers a remarkable reduction in computational cost:
V4-Pro achieves a 73% reductionin overall inference FLOPs compared to V3.2.V4-Flash achieves approximately 90% reductionin inference FLOPs.KV Cache size at 1M tokens: V4-Pro uses only 10% of V3.2’s cache; V4-Flash uses only 7%.
Compressed Sparse Attention (CSA) works by first compressing the KV cache by a factor of 4 (to 25% of its original size), then applying query-dependent top-1024 KV selection on the compressed entries. This dual-compression approach means that both the KV cache itself is dramatically smaller and the attention computation is sparse even within the compressed representation. The result is higher-resolution, query-adaptive attention with excellent balance between local detail and global context.
Heavily Compressed Attention (HCA) takes compression even further, reducing the KV cache by a factor of 128 (to less than 1% of the original size). Unlike CSA, HCA drops sparse selection entirely and runs full dense attention over the heavily compressed sequence. Because the compressed sequence is so short, dense attention becomes computationally feasible, providing strong global context at extremely low cost.
By interleaving CSA and HCA layers throughout the model, V4 achieves both fine-grained, query-dependent attention (via CSA) and broad global context (via HCA) in a single unified architecture.
The third major innovation is the Engram Conditional Memory module, introduced in a paper titled “Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models” (arXiv:2601.07372, January 2026, cited by 26 papers). Engram addresses a fundamental bottleneck in transformer architectures: the same attention mechanism is used for both factual knowledge retrieval and computational reasoning, creating a tension where improving one often degrades the other.
Engram resolves this by introducing a parallel memory channel alongside the transformer’s attention mechanism. It is grounded in classic N-gram structure but modernized with learned embeddings, providing O(1) time complexity knowledge lookup — constant-time retrieval regardless of the total size of the knowledge base. This is achieved through hash-based lookup of n-gram patterns with hash embeddings, offloading static and factual knowledge retrieval to DRAM rather than GPU compute.
The paper describes this as creating a “new axis of sparsity” that complements the existing MoE “conditional computation” axis. Together, these two axes form a complete sparsity framework:
Conditional Computation (MoE)— which neurons to activate per token.Conditional Memory (Engram)— which memory entries to retrieve per token.
The allocation law governing the trade-off between neural computation and static memory is a key finding. DeepSeek identifies a U-shaped scaling law and determines that under a fixed parameter and FLOPs budget, reallocating 20-25% of sparse capacity from experts into Engram memory optimizes overall model performance. The practical impact is significant: Engram produces 3-5 point benchmark improvements and increases Needle-in-a-Haystack accuracy from 84.2% to 97% on a 27-billion-parameter model.
DeepSeek V4’s training pipeline incorporates several advanced reinforcement learning (RL) innovations that build upon the GRPO (Group Relative Policy Optimization) algorithm first introduced with DeepSeekMath in 2024. GRPO replaces traditional PPO (Proximal Policy Optimization), which requires a separate critic/value model, with group-based statistical normalization. For each question, GRPO generates G candidate solutions (e.g., G=64), scores each via a reward function, and computes relative advantages by normalizing rewards within the group. This eliminates the need for a separate value function network, greatly reducing memory consumption and computational complexity.
