DeepSeek released its V4 model on April 24, 2026—and within 72 hours, it was running benchmarks on par with GPT-4o and Claude 3.5.
Key Takeaways
- DeepSeek V4 handles much longer prompts than its predecessor, thanks to a new architecture optimized for efficiency.
- It’s the first DeepSeek model optimized for Huawei’s Ascend chips, a signal China is preparing to operate independently of Nvidia.
- Despite being open source, V4 performs on par with closed models from OpenAI, Anthropic, and Google.
- The release intensifies scrutiny over China’s AI self-reliance amid tightening export controls and geopolitical friction.
- V4’s launch bypassed traditional fanfare—no blog post, no press release—just a GitHub update and API access.
China’s Answer to Silicon Valley Just Got Sharper
When DeepSeek dropped V4 Friday morning Beijing time, the AI community noticed immediately. Not because of any announcement, but because developers started posting side-by-side comparisons on X and Hugging Face. In those early tests, V4 matched or exceeded GPT-4o on the GSM8K math benchmark, tied Claude 3.5 on MMLU, and outperformed both on long-context reasoning tasks up to 128K tokens.
That’s not supposed to happen—not this fast, not from a company outside the US-funded AI corridor. But DeepSeek isn’t following the usual playbook. It didn’t host a keynote. It didn’t release a research paper. It just released the model weights, updated its API documentation, and let performance speak.
And the performance is 350 billion parameter strong—on par with OpenAI’s largest models, but with one critical difference: it was trained and runs efficiently on Huawei Ascend 910B chips.
The Real Test: Can China Replace Nvidia?
Nvidia’s H100 has been the gold standard for AI training since 2023. US export controls have blocked China from acquiring them in volume. That was supposed to slow Chinese AI development by 18 to 24 months. But DeepSeek V4 suggests that window is closing—fast.
V4 is the first flagship model explicitly optimized for Huawei’s Ascend architecture. That’s not just a technical footnote. It’s a strategic pivot. For years, Chinese AI firms relied on workarounds—smuggled H800s, cloud rentals, software hacks—to compensate for hardware shortages. Now, DeepSeek is betting the entire stack can be domestic.
And it’s not just DeepSeek. Alibaba, Baidu, and 4Paradigm have all announced Ascend-optimized models in the past six months. But V4 is the first to match US performance at scale. If it holds, it means China’s AI ecosystem can grow without Western chips.
Why the Ascend 910B Matters
The Ascend 910B is Huawei’s answer to the H100. It delivers about 70% of the H100’s theoretical performance on AI workloads. That gap used to be a dealbreaker. But software optimizations—especially in memory management and distributed training—have narrowed the real-world difference.
DeepSeek’s engineers reportedly spent months rearchitecting their training pipeline to minimize data movement and maximize tensor core utilization on the 910B. The result? V4 trains 15% slower than an equivalent model on H100 clusters—but at a fraction of the cost and without violating US sanctions.
- Ascend 910B availability: unrestricted within China
- Estimated cost per training run: 30–40% lower than H100 equivalent
- Domestic production capacity: growing rapidly via SMIC 7nm
- Software stack: CANN 8.0 now supports PyTorch and JAX natively
- Cloud access: Huawei Cloud offers Ascend instances in 12 Chinese provinces
Open Source as a Weapon
DeepSeek keeps its models open source. That’s unusual. OpenAI, Anthropic, Google—they’ve all gone closed. Their logic: control the model, control the moat.
DeepSeek’s bet is the opposite: openness accelerates adoption, especially in regions skeptical of US tech dominance. Developers in Southeast Asia, Africa, and Latin America are already deploying V4 on local clouds. Some are fine-tuning it for agriculture, local languages, and public health.
That’s not charity. It’s strategy. Every fine-tuned instance generates data. Every deployment builds ecosystem lock-in. And because the model is open, DeepSeek doesn’t need to operate globally to benefit from global use.
It’s the Android playbook—but for AI. US firms are stuck in a walled-garden arms race, while DeepSeek is seeding the ground.
What the US Doesn’t Understand About Chinese AI
The US assumes AI dominance is about compute scale and venture capital. But China’s model is different: state-backed infrastructure, vertical integration, and long-term tolerance for loss-making R&D.
DeepSeek doesn’t need to be profitable. It doesn’t need to answer to shareholders. It needs to deliver capability—and it’s doing that on Huawei’s hardware, with support from national labs and university partnerships.
That changes the game. US firms are racing to monetize AI before the bubble bursts. Chinese firms are racing to build foundation models that can run on domestic silicon. One is a sprint. The other is a marathon.
The Implications for Global AI Development
If V4 proves stable and scalable, it’ll force a reassessment of the global AI landscape. Right now, developers outside the US face a choice: pay premium prices for OpenAI, risk latency with EU models, or rely on underpowered open alternatives.
V4 offers a third path: high performance, no censorship, open weights, and low-cost deployment on Ascend or even repurposed gaming GPUs. That’s already attracting startups in Indonesia, Nigeria, and Chile.
But there’s a catch. Ascend optimization means V4 runs best on Huawei hardware. And Huawei is blacklisted in many Western countries. So while the model is open, the optimal stack isn’t globally accessible.
This creates a bifurcated AI ecosystem: one branch rooted in US chips and APIs, the other in Chinese silicon and open models. We’re not just seeing a tech split—we’re seeing a computational iron curtain form.
What This Means For You
If you’re a developer building AI applications, DeepSeek V4 gives you a new option—one that’s fast, open, and doesn’t require a $200/month API bill. You can run it locally, fine-tune it freely, and deploy it without rate limits. That’s huge for edge use cases, offline systems, or apps in regions with poor connectivity to US clouds.
But you’ll have to make a choice. Do you optimize for the Nvidia-OpenAI axis, or do you prepare for a world where Chinese AI stacks are first-class citizens? The hardware divide means you can’t fully optimize for both. And if your app needs long-context reasoning—like legal analysis or scientific review—V4 might be the best option available, regardless of geopolitics.
DeepSeek didn’t set out to redefine the AI cold war. But by releasing a model this capable on non-Nvidia hardware, it just did.
The Bigger Picture: Why It Matters Now
The timing of DeepSeek V4’s release is no accident. It comes amid a broader shift in how AI development is funded, regulated, and weaponized globally. The US Department of Commerce expanded its chip export restrictions in early 2025, specifically targeting AI accelerators bound for China—even those routed through third countries like Vietnam and Malaysia. The goal was to delay China’s ability to train competitive models for at least two years.
Yet DeepSeek V4 proves that time window has effectively collapsed. China isn’t just catching up. It’s building an alternative stack: SMIC’s 7nm process produces enough Ascend 910B chips to power thousands of training clusters. Huawei’s CANN 8.0 software stack now supports major ML frameworks. And national initiatives like the Beijing Academy of Artificial Intelligence are pooling datasets and compute to accelerate research.
This isn’t just about one company’s success. It’s about systemic resilience. While US firms depend on TSMC for chips and Amazon or Microsoft for cloud infrastructure, China is integrating everything domestically. The Ministry of Industry and Information Technology (MIIT) has mandated that all state-funded AI projects use domestic hardware by 2027. That creates guaranteed demand for models like V4.
The consequences ripple outward. Countries like Egypt, Pakistan, and Venezuela—already using Huawei for telecom infrastructure—are now evaluating Ascend-based AI for public services. They don’t need access to US clouds. They need functional, affordable models that work on available hardware. DeepSeek V4 delivers that.
This isn’t a temporary workaround. It’s the foundation of a parallel AI economy—one that operates outside US financial and technological controls.
Competing Visions: How Other Chinese Firms Are Responding
DeepSeek isn’t alone in pushing for hardware-software co-design on domestic chips. Alibaba has launched Qwen-Max-Ascend, a version of its flagship model optimized for 910B clusters, and is deploying it across its cloud services in Chengdu and Shenzhen. Baidu’s Ernie Bot 5.0, released in February 2026, achieved 88% of GPT-4o’s MMLU score while running entirely on Ascend chips. Even Tencent has shifted its AI roadmap, quietly retiring its reliance on imported A100s in favor of hybrid Ascend-gaming GPU clusters.
What sets DeepSeek apart is its openness. While Alibaba and Baidu keep their optimized models behind API walls, DeepSeek released V4’s weights under the MIT License. That invites scrutiny, yes—but also rapid iteration. Within 48 hours of the release, over 1,200 forks appeared on Hugging Face, including versions tuned for low-memory devices and medical diagnostics.
Meanwhile, US firms are moving in the opposite direction. OpenAI hasn’t open-sourced a major model since GPT-2. Google restricts access to Gemini Ultra. Anthropic sells usage through closed enterprise contracts. Their bet is on control and monetization. DeepSeek’s bet is on reach and influence.
The contrast couldn’t be starker. In the US, AI is a product. In China, increasingly, it’s infrastructure. And infrastructure gets funded, protected, and scaled—regardless of quarterly returns.
The Hardware-Software Feedback Loop
What makes V4 truly significant isn’t just that it runs on Ascend chips. It’s that the model’s architecture was co-designed with the hardware. That’s a shift from the past, where Chinese firms simply ported Western model designs onto available silicon, accepting performance penalties.
DeepSeek’s engineers used Huawei’s CANN profiling tools to identify bottlenecks in attention layers and feed-forward networks. They redesigned the model’s memory layout to reduce off-chip data transfers—a major drag on 910B performance. They also implemented sparse activation patterns that align with the Ascend’s tensor cores, boosting throughput by 22% in long-sequence tasks.
This kind of deep integration mirrors what Nvidia achieved with CUDA and Transformer engines in H100 clusters. But where Nvidia controls both hardware and software, Huawei and DeepSeek are building their own version through partnership, not ownership. No single entity calls the shots. Instead, a network of state labs, chipmakers, and startups collaborate under national strategic goals.
The feedback loop is now self-sustaining: better software improves model performance on domestic chips, which justifies more investment in chip fabrication, which in turn enables larger, more efficient models. It’s a cycle that doesn’t depend on external inputs. And once it reaches critical mass, it becomes very hard to disrupt.
Western observers often miss this dynamic. They focus on raw FLOPS or benchmark scores. But resilience, accessibility, and ecosystem coherence are just as important in the long run. DeepSeek V4 isn’t just a model. It’s a milestone in a decade-long effort to build an autonomous AI stack—and it shows that China’s plan is working.
Can an open-source model trained on domestic chips become the default for the Global South—and challenge US AI supremacy on performance alone?
Sources: MIT Tech Review, original report
