The Download: DeepSeek’s latest AI breakthrough, and the race to build world models

```json { "title": "DeepSeek V4 Launches With Huawei Chips and 1M-Token Context", "metaDescription": "DeepSeek V4 debuts with 1.6 trillion parameters, Huawei Ascend chip support, and prices undercutting GPT-5.5 by a wide margin. Here's what it means.", "content": "<h2>DeepSeek V4 Is Here — and It's Built for China's Chip Ecosystem</h2><p>On April 24, 2026, Chinese AI startup DeepSeek released a preview of its long-awaited V4 flagship model, marking the company's most significant release since R1 in January 2025. The launch arrives amid an intensifying global AI race, a freshly released GPT-5.5 from OpenAI, and a White House accusation that Chinese entities are conducting industrial-scale campaigns to distill capabilities from US AI systems. DeepSeek V4 is notable not just for its technical specs — though those are substantial — but for what it signals about China's push toward AI self-sufficiency built on domestic hardware.</p><h2>Two Models, One Million Tokens, and a Fraction of the Price</h2><p>The V4 series ships in two variants, both open-source under the MIT license. <strong>DeepSeek-V4-Pro</strong> carries 1.6 trillion total parameters with 49 billion activated per token, while <strong>DeepSeek-V4-Flash</strong> has 284 billion total parameters with 13 billion activated. Both models were pre-trained on more than 32 trillion diverse, high-quality tokens and support a one-million-token context window — a leap enabled by a novel Hybrid Attention Architecture that combines Compressed Sparse Attention and Heavily Compressed Attention mechanisms.</p><p>According to DeepSeek's technical documentation on Hugging Face, V4-Pro requires only 27% of the single-token inference FLOPs and 10% of the KV cache compared with DeepSeek V3.2 in the one-million-token context setting. The hybrid attention mechanism reduces memory requirements by a factor of 9.5 to 13.7 compared to V3.2 under the same conditions — a meaningful engineering achievement that makes the extended context window practically usable rather than just theoretically possible.</p><p>On pricing, the gap between DeepSeek and US competitors is stark. V4-Flash is priced at $0.14 per million input tokens and $0.28 per million output tokens, making it the cheapest of the small frontier models currently available. V4-Pro comes in at $1.74 per million input tokens and $3.48 per million output tokens. OpenAI's GPT-5.5, which launched one day before DeepSeek V4, is priced at $5 per million input tokens and $30 per million output tokens — making V4-Pro roughly three times cheaper on input and nearly nine times cheaper on output.</p><p>DeepSeek also claims strong benchmark performance. In its own technical documentation, the company states: <em>"In world knowledge benchmarks, DeepSeek V4-Pro significantly leads other open source models and is only slightly outperformed by the top-tier closed-source model Gemini-3.1-Pro."</em> Independent verification of these benchmarks is ongoing, and readers should weigh self-reported performance claims accordingly.</p><h2>The Huawei Factor: Domestic Chips Take Center Stage</h2><p>Perhaps the most geopolitically significant aspect of V4 is its integration with Huawei's domestic Ascend AI chips. According to MIT Technology Review, V4 is DeepSeek's first model explicitly optimized for domestic Chinese chips. DeepSeek reportedly gave early prerelease access only to Chinese chipmakers, bypassing Nvidia and AMD entirely.</p><p>Huawei confirmed its involvement hours after the preview dropped, stating: <em>"Through close technical collaboration … the entire Ascend supernode product line now supports the DeepSeek-V4 series models."</em> Reuters reported that Huawei's chips were used for part of V4-Flash's training and that V4 is fully supported on Huawei's Ascend 950-based supernode clusters, which combine large clusters of Ascend 950 chips into what the company calls its 'Supernode' architecture.</p><p>According to Bloomberg, the delayed V4 release was attributed to a strategic shift toward deeper integration with China's domestic chip ecosystem, with DeepSeek spending months reworking its software stack to optimize performance on Huawei's Ascend chips. This stands in contrast to earlier DeepSeek models, including R1 and V3, which were trained on Nvidia hardware. The move is widely read as a deliberate step toward reducing China's AI sector dependence on US semiconductor supply chains.</p><p>The timing was pointed. One day before V4's release, White House director of the office of science and technology policy Michael Kratsios accused foreign entities primarily based in China of conducting 'industrial-scale' campaigns to 'distill' frontier AI models from US companies — essentially building on the capabilities of models like GPT-4 without paying for the underlying research investment.</p><h2>Market and Competitive Context</h2><p>V4's reception in financial markets was notably quieter than R1's January 2025 debut, which triggered significant sell-offs in US semiconductor stocks. According to CNBC, shares of Chinese AI rivals MiniMax and Zhipu each fell around 8% on the Hong Kong market on the day of V4's release, suggesting the primary competitive pressure from this launch falls on domestic Chinese players rather than US firms.</p><p>Since R1's release, DeepSeek has faced intensifying competition within China's own AI sector, with players including Alibaba and ByteDance releasing their own models. V4 lands into that crowded domestic landscape as much as it does into the global frontier model race.</p><p>On Hugging Face, V4 quickly became the top trending model on the platform. According to TechWireAsia, Lewis Tunstall, a machine learning engineer at Hugging Face, noted it was the fastest model to reach that position.</p><h2>Expert Reactions</h2><p>Analysts have been measured but attentive in their assessments of V4's significance.</p><p>Wei Sun, principal analyst at Counterpoint Research, highlighted the chip angle as the story's most consequential dimension: <em>"It allows AI systems to be built and deployed without relying solely on Nvidia, which is why V4 could ultimately have an even bigger impact than R1 — accelerating adoption domestically and contributing to faster global AI development overall."</em></p><p>Neil Shah, vice president of research at Counterpoint Research, offered a more direct characterization: <em>"DeepSeek's V4 preview is a serious flex."</em></p><p>Ivan Su, senior equity analyst at Morningstar, was more cautious about market impact, noting that V4's debut is unlikely to replicate R1's shock effect because traders have already adjusted their expectations around Chinese AI competitiveness. He added context on how the competitive framing itself has shifted: <em>"This is a framing that didn't exist with R1, and that alone tells you how much domestic competition has intensified."</em></p><h2>The Bigger Picture: World Models and the Next AI Frontier</h2><p>DeepSeek V4 is not the only major AI development shaping the landscape this week. MIT Technology Review has been tracking a parallel race to build what researchers call 'world models' — AI systems capable of understanding and simulating physical reality, rather than simply predicting text.</p><p>Key players investing in this emerging paradigm include Google DeepMind, Stanford professor Fei-Fei Li's World Labs, and Yann LeCun, who has departed Meta to form a world-model-focused startup. OpenAI is also entering the space by reallocating resources from its shuttered Sora video app toward what it describes as 'longer-term world simulation research.'</p><p>One of the more concrete real-world applications of world model technology is being developed by Niantic Spatial, the AI spinout from Pokémon Go maker Niantic. The company is training a world model using 30 billion images of urban landmarks crowdsourced from Pokémon Go players, and has partnered with Coco Robotics to use the technology to help delivery robots navigate cities with centimeter-level accuracy. Brian McClendon, CTO at Niantic Spatial, offered some perspective on the scale of the underlying data source: <em>"Five hundred million people installed that app in 60 days."</em> That global install base, accumulated over years of gameplay, has quietly become a training dataset for physical-world AI.</p><p>The convergence of cheaper, more capable language models like DeepSeek V4 and the emerging generation of world models points toward AI systems that are simultaneously more affordable to run and more grounded in physical reality — a combination with significant implications for robotics, logistics, healthcare, and productivity tooling.</p><h2>What's Next</h2><p>V4 is currently a preview release, and full model weights and additional technical details are expected to follow. The extent to which DeepSeek's Huawei chip optimization translates into production-scale performance — and whether other Chinese AI labs will follow a similar hardware strategy — remains to be seen.</p><p>The broader question of whether China can build a genuinely self-sufficient AI infrastructure, from chips to models to deployment, is one that V4 advances but does not resolve. DeepSeek's decision to bypass Nvidia and AMD in its prerelease access suggests the company is treating domestic chip optimization as a strategic priority, not just a technical footnote.</p><p>On the world models front, the competitive landscape is moving quickly, with multiple well-funded teams pursuing different architectural approaches. Whether any single approach achieves the kind of general physical reasoning that the field is targeting remains an open question.</p><p>For more tech news, visit our <a href="/news">news section</a>.</p><h2>Why This Matters for Your Productivity</h2><p>Cheaper, more capable AI models don't just reshape corporate strategy — they change what's accessible to individuals. DeepSeek V4-Flash at $0.14 per million input tokens means that AI-assisted research, writing, analysis, and decision-support tools are becoming dramatically more affordable to build and use. As world models mature and language models get cheaper, the tools available to support your focus, health decisions, and daily productivity will grow more powerful and more personalized. Staying informed about these shifts puts you in a position to adopt the right tools at the right time. <a href="/#waitlist">Join the Moccet waitlist to stay ahead of the curve.</a></p>", "excerpt": "DeepSeek released a preview of its V4 flagship model on April 24, 2026, shipping in two open-source variants with up to 1.6 trillion parameters, a one-million-token context window, and pricing that undercuts GPT-5.5 by a significant margin. The release is notable for its first-of-its-kind optimization for Huawei's domestic Ascend chips, signaling a strategic shift toward Chinese AI self-sufficiency. It arrives alongside a growing industry race to build world models capable of simulating physical reality.", "keywords": ["DeepSeek V4", "Huawei Ascend chips", "AI world models", "Chinese AI", "large language models"], "slug": "deepseek-v4-huawei-chips-world-models-2026" } ```