China's DeepSeek releases preview of long-awaited V4 model as AI race intensifies

```json { "title": "DeepSeek V4 Preview Released: 1.6T Parameters, 1M Token Context", "metaDescription": "DeepSeek releases its V4 preview on April 24, 2026, featuring two open-source MoE models with one million token context windows and novel AI architecture.", "content": "<h2>DeepSeek Launches V4 Preview With Two Open-Source Models as Global AI Competition Heats Up</h2>\n\n<p>Chinese AI startup DeepSeek released a preview version of its long-awaited V4 large language model on April 24, 2026, unveiling two distinct Mixture-of-Experts (MoE) models under an open-source MIT license. The release marks one of the most anticipated AI launches of 2026 — and arrived on the same day that OpenAI shipped GPT-5.5, underscoring just how fierce the global AI race has become.</p>\n\n<p>The two models — <strong>DeepSeek-V4-Pro</strong> and <strong>DeepSeek-V4-Flash</strong> — are available immediately via DeepSeek's web platform, mobile app, and API, according to the company's official website updated April 24, 2026. Both models support a context window of one million tokens, a capability DeepSeek says positions them at the frontier of long-context AI performance.</p>\n\n<h2>What's Inside DeepSeek V4: Architecture, Scale, and Efficiency</h2>\n\n<p>According to DeepSeek's official Hugging Face model cards, DeepSeek-V4-Pro is a 1.6 trillion total parameter model, with 49 billion parameters activated per token during inference. DeepSeek-V4-Flash is the smaller and more efficient variant, carrying 284 billion total parameters with 13 billion activated per token. Both models were pre-trained on more than 32 trillion diverse, high-quality tokens, followed by a comprehensive post-training pipeline.</p>\n\n<p>The efficiency gains in V4-Pro are notable. In the one-million-token context setting, DeepSeek-V4-Pro requires only 27% of the single-token inference FLOPs and just 10% of the KV cache compared with DeepSeek-V3.2, according to the company's own model card. That kind of compute reduction at extreme context lengths is a meaningful engineering achievement, and one that has direct implications for the cost of deploying frontier AI at scale.</p>\n\n<p>Architecturally, the V4 series introduces a hybrid attention system combining <strong>Compressed Sparse Attention (CSA)</strong> and <strong>Heavily Compressed Attention (HCA)</strong>, designed to improve efficiency at long context lengths. The models also incorporate <strong>Manifold-Constrained Hyper-Connections (mHC)</strong>, a technique intended to strengthen residual connections in the network. Pre-training used FP4 and FP8 mixed precision — with MoE expert parameters at FP4 and most other parameters at FP8 — according to DeepSeek's official model card.</p>\n\n<p>DeepSeek-V4-Flash is positioned as the accessible option in the lineup. According to AFP reporting, the company described it as "a more efficient and economical choice" due to its smaller parameter count, making it suitable for use cases where speed and cost matter more than maximum capability.</p>\n\n<h2>Benchmark Performance: Where V4 Stands Against Rivals</h2>\n\n<p>DeepSeek described V4 as having achieved "leadership in both domestic and open-source fields across agent capabilities, world knowledge, and reasoning performance," according to the company's official social media statement reported by AFP. The company also stated that "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, (Google's) Gemini-Pro-3.1."</p>\n\n<p>On coding benchmarks, the numbers are striking. According to a Kingy.ai analysis of DeepSeek's official Hugging Face model card, DeepSeek-V4-Pro-Max achieved a Codeforces rating of 3206, ahead of GPT-5.4 xHigh at 3168 in DeepSeek's own benchmark table. On the SWE Verified benchmark — a measure of real-world software engineering capability — DeepSeek-V4-Pro-Max scored 80.6, tying leading models in that category.</p>\n\n<p>It is worth noting that these benchmark figures come from DeepSeek's own model card and third-party analyses of that card. Independent, peer-reviewed evaluations of V4 have not yet been published as of April 24, 2026, and the preview status of this release means performance characteristics may shift before a full production launch.</p>\n\n<h2>The Huawei Hardware Question and the Geopolitical Undercurrent</h2>\n\n<p>The V4 release carries significance well beyond benchmark scores. According to Reuters, citing The Information, DeepSeek did not give Nvidia or AMD early access to V4 for optimization — while domestic Chinese suppliers including Huawei did receive early access. Reuters also reported, as of early April 2026, that DeepSeek V4 was expected to run on Huawei's latest chips.</p>\n\n<p>This hardware dimension matters enormously in the context of ongoing U.S. semiconductor export restrictions on China. If DeepSeek has trained and optimized V4 on Huawei's Ascend silicon rather than Nvidia hardware, it would represent a concrete demonstration that Chinese AI development can continue to advance despite export controls — a point that carries implications far beyond any single model release.</p>\n\n<p>Wei Sun, principal AI analyst at Counterpoint Research, put the broader significance plainly: <em>"It's important to know because at one level, it is a signal of China's AI self-sufficiency trajectory."</em></p>\n\n<p>Stephen Wu, founder of the Carthage Capital fund, had anticipated exactly this possibility ahead of the release: <em>"if they have successfully trained V4 entirely on Huawei silicon, it signals a material shift in the geopolitical tech landscape."</em></p>\n\n<p>DeepSeek has not publicly confirmed the specific hardware configuration used to train V4 as of the preview release date. Some unverified prior reporting alleged that V4 may have been trained on smuggled Nvidia Blackwell chips — a claim that Nvidia reportedly described as appearing "farfetched," according to The Information. The MIT-licensed release of model weights does not, on its own, resolve the hardware question.</p>\n\n<h2>Context: DeepSeek's Rapid Rise and What V4 Represents</h2>\n\n<p>DeepSeek began in 2023 as a side project of a hedge fund that had access to a cache of powerful Nvidia processors, according to the Hong Kong Free Press. It entered the global AI conversation in dramatic fashion in January 2025, when its R1 deep-reasoning model — built on the V3 model released in December 2024 — matched leading U.S. AI systems at a fraction of the reported training cost. The release sent U.S. tech shares tumbling and prompted President Donald Trump to call it a "wake-up call" for American firms.</p>\n\n<p>Since then, DeepSeek has become a widely used platform in China and in emerging markets including Southeast Asia and the Middle East. V4 represents the company's next step: a model that maintains the open-source, cost-efficient ethos that made R1 disruptive, while substantially expanding context window capability and, according to DeepSeek's own claims, pushing performance closer to the frontier of closed-source systems.</p>\n\n<p>The timing of the release adds to the competitive intensity. OpenAI also shipped GPT-5.5 on April 24, 2026, the same day as the DeepSeek-V4 preview, according to ofox.ai. The simultaneous releases — one from a Chinese open-source startup, one from the dominant U.S. AI lab — encapsulate the state of the industry heading into mid-2026.</p>\n\n<p>Wu had framed the stakes of V4 ahead of launch: <em>"I expect the upcoming DeepSeek V4 release will not just be a software update; it will be a highly capable, open-source model that handles massive context windows at a fraction of the cost."</em> The preview release appears consistent with that framing, though full production performance and independent evaluations remain to be seen.</p>\n\n<h2>What Comes Next for DeepSeek V4</h2>\n\n<p>The April 24 release is explicitly a preview, meaning DeepSeek has not yet committed to a final production launch date or confirmed that all features are stable. The model's official paper is titled <em>DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence</em>, and the technical depth described in the model cards suggests a full research publication will accompany or follow the final release.</p>\n\n<p>For developers and researchers, the immediate availability of model weights under the MIT License — one of the most permissive open-source licenses — means that independent evaluation, fine-tuning, and deployment can begin immediately. That openness has been central to DeepSeek's strategy and is a key differentiator from closed-source competitors.</p>\n\n<p>The hardware question — specifically, which chips were used in training and whether V4 can be efficiently served on Huawei's Ascend infrastructure at scale — is likely to draw continued scrutiny from analysts, policymakers, and competitors alike. The answer has implications not just for DeepSeek, but for how the U.S. government and industry assess the effectiveness of semiconductor export controls as a geopolitical tool.</p>\n\n<p>For now, DeepSeek-V4 Preview is live. Whether the full release matches the preview's promise, and what the hardware story ultimately reveals, are the two questions worth watching most closely.</p>\n\n<p>For more tech news, visit our <a href=\"/news\">news section</a>.</p>\n\n<h2>Why This Matters for Your Productivity</h2>\n\n<p>The rapid advance of open-source AI models like DeepSeek V4 — with one-million-token context windows and drastically lower inference costs — is accelerating the pace at which powerful AI tools become accessible to individuals and teams, not just large enterprises. For anyone working to optimize how they think, work, and perform, staying informed about these shifts is no longer optional. At Moccet, we track the intersection of technology and human performance so you don't have to do it alone. Join the <a href=\"/#waitlist\">Moccet waitlist</a> to stay ahead of the curve.</p>", "excerpt": "DeepSeek released a preview of its long-awaited V4 model on April 24, 2026, comprising two open-source Mixture-of-Experts models — V4-Pro with 1.6 trillion parameters and V4-Flash with 284 billion — both supporting one-million-token context windows. The release, which landed on the same day as OpenAI's GPT-5.5, marks a significant moment in the accelerating global AI race and raises fresh questions about China's AI hardware self-sufficiency.", "keywords": ["DeepSeek V4", "DeepSeek-V4-Pro", "open-source AI", "large language model", "China AI"], "slug": "deepseek-v4-preview-released-1-6t-parameters-1m-token-context" } ```