DeepSeek V4 Comprehensive Analysis: The Trillion-Parameter Revolution, Ascend Adaptation, and NVIDIA’s Crossroads

Introduction

On April 24, 2026, just hours after OpenAI released GPT-5.5, the much-anticipated DeepSeek-V4 preview was officially launched and simultaneously open-sourced. The V4 series includes two MoE (Mixture of Experts) models — DeepSeek-V4-Pro and DeepSeek-V4-Flash — both natively supporting a 1-million-token context window and fully open-sourced under the MIT license.

If the breakthrough of DeepSeek-V3 and R1 was achieving performance comparable to closed-source models at an ultra-low training cost, the significance of V4 runs even deeper. It not only achieves a paradigm shift at the architectural level — from “parameter competition” to “efficiency first” — but also, through its full adaptation to Huawei’s Ascend domestic computing platform, drops a heavy question mark on the global AI industry across three dimensions: technology, business, and geopolitics.

01 Technical Specifications: Hard Power Behind the Numbers

1.1 Parameter Scale and Model Matrix

The DeepSeek-V4 series includes two versions:

V4-Pro is positioned as the flagship model for ultimate performance and deep reasoning, corresponding to the “expert mode” on the official website; V4-Flash focuses on high speed and high efficiency, corresponding to “fast mode”.

1.2 Architectural Innovation: Hybrid Attention Mechanism

V4’s core technical breakthrough lies in its novel attention mechanism design. In traditional Transformers, attention complexity grows quadratically with sequence length, making a 1‑million-token context window nearly impossible to implement in engineering. DeepSeek-V4 adopts a hybrid attention architecture called “Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA),” performing hierarchical compression and retrieval at the token level.

Specifically, CSA uses dynamic sequence compression to reduce KV cache entries, then applies DeepSeek Sparse Attention (DSA) to sparsify the attention matrix, thereby lowering computational overhead. HCA employs an even more aggressive compression strategy, merging KV entries of multiple tokens into a single compressed entry, significantly reducing KV cache size.

At the MoE architecture level, V4 adopts a “Mega Kernel” design with 384 experts per layer, activating 6 experts per inference, continuing DeepSeek’s leadership in MoE. The training backend also introduces several new techniques, including the Muon optimizer and a derived reinforcement learning solution with GRPO and KL divergence correction.

1.3 Efficiency Revolution: Plunging Compute Demand

The direct result of architectural innovation is a revolutionary leap in computational efficiency. Under the 1‑million‑token context setting, V4-Pro’s single-token inference FLOPs are only 27% of V3.2’s, and KV cache size is compressed to 10% of its predecessor; V4-Flash pushes these numbers down to 10% and 7%, respectively. In other words, while the context window expands from 128K to 1M (a theoretical 8× increase), the per-token compute demand drops sharply — a fundamental breakthrough enabled by the hybrid attention mechanism.

1.4 Performance Evaluation: A “Follower” Closing in on Closed Source

Regarding performance evaluation, DeepSeek has remained relatively restrained in its official stance. The official report clearly states that V4’s capabilities still lag behind GPT-5.4 and Gemini-3.1-Pro, with the overall development trajectory lagging behind state-of-the-art closed-source models by about 3 to 6 months. On specific benchmarks, V4-Pro-Max tops two hard‑core tasks: Apex Shortlist (90.2%) and Codeforces Rating (3206). In mathematics, STEM, and competitive coding evaluations, it surpasses all open‑source models with published results, achieving performance comparable to the world’s best closed‑source models. In agent capabilities, V4-Pro achieves the best open‑source level in Agentic Coding evaluation, with a user experience better than Claude Sonnet 4.5 and delivery quality close to Opus 4.6’s non‑thinking mode.

02 Comprehensive Parameter Comparison with V3/R1

To better understand V4’s generational leap, a systematic parameter and performance comparison with its predecessor is necessary.

Note: V3 data based on DeepSeek-V3.2 official technical report and industry disclosures

The table clearly shows that while V4’s total parameters have ballooned to 1.6 trillion — making it one of the largest open‑source MoE models in terms of total parameters — its active parameters increased by only about 32% compared to V3.2. This means V4’s design is not about simply stacking parameters, but rather expanding the breadth and depth of the expert pool to increase the model’s “knowledge reserve,” while relying on more sophisticated MoE routing to keep computational load under control. What truly signals V4’s generational leap is the 8× expansion of the context window from 128K to 1M, along with the accompanying efficiency revolution — the dramatic drop in FLOPs and KV cache clearly shows that the architectural innovation in attention is V4’s true soul.

03 Full Adaptation to Huawei Ascend: Three Dimensions of Historic Significance

One of the most closely watched aspects of DeepSeek-V4 is its comprehensive adaptation and collaboration with Huawei’s Ascend domestic computing platform. The depth and breadth of this collaboration far exceed any previous linkage between a model and domestic hardware.

3.1 Full‑Stack Adaptation for Training and Inference

V4 was designed from the outset with native operator optimization for the Ascend platform, rather than post‑hoc migration and adaptation. The Ascend 910B cluster already supports the training of DeepSeek’s entire model series, and Huawei Computing has announced that the entire Ascend SuperPod product line will simultaneously support the deployment of DeepSeek V4 series models. After deep optimization, DeepSeek-V4’s inference speed on Huawei Ascend chips has increased 35‑fold compared to early versions, and Huawei’s self‑developed CANN (Compute Architecture for Neural Networks) framework, through dynamic graph compilation optimization, can further improve the computational efficiency of DeepSeek’s attention mechanism by 23%.

From a strategic perspective, the core positioning of the DeepSeek‑Huawei collaboration is “domestic large model + domestic computing power deeply integrated to create a self‑controllable AI closed loop.” According to the collaboration timeline, the V4 flagship version is prioritized for the Ascend platform; historical/existing models still use NVIDIA compute; the future roadmap is “domestic first, dual track.”

3.2 Systemic Support for Domestic Substitution

Adapting to Huawei Ascend carries significance far beyond the migration of a single model. By December 2025, the Huawei Ascend ecosystem had formed a landscape of “diverse hardware, full‑stack open‑source software, large‑scale developers, and deep industry adoption” — with over 6.65 million developers, more than 80 hardware partners, and over 23,900 industry‑certified solutions. The Ascend 910B chip delivers 320 TOPS @FP16, with measured performance reaching 91% of an NVIDIA A100 cluster.

At the cluster level, Huawei has released the Atlas 950 SuperPoD and Atlas 960 SuperPoD supernode products, supporting 8,192 and 15,488 Ascend cards respectively, with compute scales exceeding 500,000 cards and reaching million‑card levels. Such clusters provide a scalably expandable domestic computing foundation for training and inference of hundred‑billion to trillion‑parameter models.

More importantly, during the adaptation from CUDA to CANN, the biggest challenge the DeepSeek engineering team overcame was not operator rewriting itself, but precision alignment — ensuring that the same model produces mathematically identical results on NVIDIA and Ascend platforms. This places extremely high demands on the underlying consistency of the compiler and numerical computation precision. Overcoming this essentially sets a technical benchmark and engineering paradigm for subsequent domestic adaptation of more large models.

3.3 The “Price Slasher’s” Compute Gap

DeepSeek’s enduring “price slasher” image faces a new challenge with V4. The API pricing for V4-Pro is: 1 RMB per million tokens for cache‑hit input, 12 RMB per million tokens for cache‑miss input, and 24 RMB per million tokens for output. This represents a roughly 6‑8 fold increase over the ultra‑low pricing of the V3.2 era.

However, the price increase is not due to DeepSeek abandoning its cost‑effectiveness approach, but rather a bottleneck in high‑end compute supply. DeepSeek officially stated, “Due to limited high‑end compute, the Pro service’s throughput is currently very constrained. It is expected that after the mass launch of Ascend 950 SuperPods in the second half of the year, the Pro price will be significantly reduced.” In other words, the current price reflects “insufficient compute capacity,” not a change in pricing strategy. The mass launch of Ascend 950 SuperPods will be the key turning point for V4-Pro’s price reduction and an important yardstick for measuring the maturity of the domestic compute supply chain.

04 Impact on NVIDIA: Calm Before the Storm

The potential impact of DeepSeek-V4 on global AI chip giant NVIDIA needs to be viewed from a broader perspective — not just V4’s technical specifications, but the slow but profound shift in the power structure of the entire AI industry.

4.1 Indirect Erosion of Performance‑for‑Price

When the predecessor DeepSeek-R1 was released, the market experienced a breathtaking shock — on what some traders called “AI Black Monday” in January 2025, NVIDIA’s stock price plunged 17% intraday, erasing nearly $600 billion in market capitalization. After V4’s release, market reaction has been relatively calm, but this does not mean the intensity of the shock has diminished — only that its form has changed.

V4’s most fundamental threat to NVIDIA is this: it dramatically reduces compute demand for long‑context inference scenarios through software‑side architectural innovation — single‑token FLOPs down 73%, KV cache compressed 90%. This means that, given the same amount of compute, the same inference task can serve more users, or the same quality of service can be achieved with less compute. In the past, long‑context and agent‑like applications were considered the exclusive domain of high‑end GPUs, because their memory and bandwidth requirements were far beyond what cheap hardware could bear. By innovating the attention mechanism itself, V4 substantially lowers the depth of that moat, thereby logically weakening NVIDIA’s core advantage in inference for high‑end GPUs.

4.2 The Real Threat of a Domestic Replacement Ecosystem

Even more alarming for NVIDIA is the demonstration effect and ecosystem‑leveraging capability of V4’s collaboration with Ascend. At the end of 2023, the gap between top Chinese and US large models across various dimensions was still 20%‑30%. By April 2026, Stanford University’s HAI Lab’s AI Index Report showed that the performance gap between Chinese and US large models had narrowed to 2.7%, essentially achieving technological parity. In this process, DeepSeek played a crucial “bridge” role — actively adapting and empowering domestic hardware through software innovation, paving the way for domestic chips.

NVIDIA CEO Jensen Huang is acutely aware of this. In an interview, he said that restricting exports of compute chips to China may indeed temporarily slow the pace of China’s AI development, but in the long run, “it will only force China to form its own ecosystem.” He further warned that if more and more open‑source models run on domestically produced compute chips, NVIDIA, even if it remains the market leader, will no longer be the only choice for AI computing.

4.3 Double Squeeze from Export Controls and Market Size

China’s strategic importance to NVIDIA is self‑evident — accounting for about 14% of its revenue, nearly $17 billion in size. However, NVIDIA’s share of China’s AI chip market has fallen from about 95% in 2022 to about 50%, as competitor Huawei Ascend catches up. Making matters worse for NVIDIA is the US export control policy — on one hand, it restricts exports of high‑end chips to China; on the other, it forces Chinese companies to accelerate domestic substitution, inadvertently speeding up the maturation of the Ascend ecosystem.

A dramatic development occurred at the end of 2025: US President Donald Trump approved exports of NVIDIA H200 chips to China, on the grounds that Huawei already provides AI computing systems with comparable performance, making continued blockade of the H200 strategically much less valuable. This policy relaxation itself is a strong signal: US policymakers have realized that export controls alone can no longer constrain the rise of Chinese AI computing.

NVIDIA’s response is two‑pronged: on one hand, it is launching China‑specific Blackwell architecture chips priced at $6,500–$8,000, trying to compete on price; on the other hand, the NVIDIA official blog published a technical article on the same day as V4’s release, showcasing inference performance of 150+ tokens/sec for DeepSeek-V4-Pro deployed on GB200 NVL72, thereby demonstrating its own platform’s ability to handle trillion‑parameter models. However, the effectiveness of this dual strategy — “technology competition plus price competition” — in slowing down the tide of domestic substitution by Huawei Ascend remains unknown.

05 V4’s True Significance: Architectural Revolution, Not Capability Domination

If we place DeepSeek-V4 in the broader coordinate system of AI industry evolution, its true historic significance is not about crushing GPT or Gemini in capability, but about a paradigm shift.

As DeepSeek itself says, the V4 preview is positioned as “reconstructing long‑context cost and laying the infrastructure for the next phase of test‑time scaling and long‑horizon tasks.” With the explosion of agent‑like applications — products like OpenClaw and Hermes — long‑context memory capability has become the core competitive factor on the watershed from “toy” to “tool.” In this new paradigm, whoever can minimize the compute cost of long contexts while maintaining model performance will seize the initiative in the next phase of commercial competition.

From this perspective, DeepSeek-V4’s breakthrough is no smaller than V3’s or R1’s — only the form of breakthrough has changed, from a cost revolution on the training side to a long‑context cost reconstruction on the inference side.

Conclusion

The release of DeepSeek-V4, combined with the full adaptation to Huawei’s Ascend ecosystem, is pushing the global AI industry toward a new distributed competitive landscape: the United States maintains leadership in process technology and per‑card compute; China leverages architectural innovation and ecosystem integration to close the gap, while also playing a unique advantage in cost control and software efficiency.

Whether NVIDIA’s “money‑printing machine” model will be challenged is not a question of “yes or no,” but “when and to what degree.” What is certain is that as open‑source models converge toward closed‑source capabilities, domestic compute performance catches up to NVIDIA, and AI application scenarios tilt toward inference, NVIDIA’s dominance is being eroded by multiple factors in concert. But as evidenced by the fact that players like OpenAI are still buying NVIDIA chips in large quantities, this erosion will be a gradual process, not an overnight upheaval.

For DeepSeek, V4 demonstrates a clear path: using open source as a banner, efficiency innovation as a core competitive edge, and domestic computing power as strategic depth, it has secured its position ahead of the turning point where the focus of the global AI race shifts from “competition in training compute” to “competition in inference cost.”

For Huawei Ascend, V4’s adaptation proves that the CANN ecosystem has reached the engineering capability to carry trillion‑parameter models — this is no longer a slogan from a “catch‑up” phase, but a realized fact.

For NVIDIA, this may not be a collapsing wall, but it has clearly seen the cracks in the wall.