DeepSeek V4 Launch Imminent: 1 Trillion Parameters, Huawei Chips, Open Weights

DeepSeek is preparing to release V4, a roughly 1 trillion parameter AI model with fully open weights that runs on Huawei's latest chips rather than Nvidia hardware. If it performs as reported, V4 would mark the most significant demonstration yet that competitive frontier AI can be built outside the Nvidia-dominated US chip ecosystem — with major implications for both US export controls and the global AI race.

What DeepSeek V4 is

According to TechNode, V4 is:

• A ~1 trillion parameter model — placing it in the same size range as the largest US frontier models
• Built using a Mixture-of-Experts (MoE) architecture — a design that activates only a subset of parameters for any given input, making very large models more computationally efficient to run
• Fully multimodal — able to process and generate text, images, and likely other data types
• Released with open weights, meaning the underlying model parameters will be publicly downloadable and deployable without DeepSeek's involvement
• Trained on Huawei Ascend chips rather than Nvidia A100 or H100 GPUs, which China cannot legally import due to US export controls

Why the Huawei chip fact matters

US export controls on advanced AI chips to China were designed with a specific theory: that limiting access to Nvidia's top-end hardware would constrain China's ability to train frontier AI models. That theory depends on there being no viable alternative to Nvidia hardware at the frontier.

DeepSeek has been systematically testing that assumption. DeepSeek V3, released in late 2025, achieved performance competitive with leading US models at a fraction of the reported training cost. If V4 achieves comparable results on Huawei hardware, it will demonstrate that the Nvidia chip dependency is not as absolute as US policymakers assumed.

The open weights factor

DeepSeek releases its models with open weights — meaning the trained model parameters are publicly available. This is in contrast to closed models like GPT-4, Claude, or Gemini, where the underlying model is only accessible through APIs.

Open weights have significant implications:

• No access controls. Any individual, organization, or government can download and run V4 without any relationship with DeepSeek or any usage restrictions DeepSeek might impose.
• Fine-tuning. Open weights can be modified and adapted. A V4 with open weights is not just a powerful model — it is a foundation for building specialized models in every domain, without ongoing cost or API dependence.
• Geopolitical spread. Open weights cross borders instantly. Export controls on chips affect hardware. They do not affect model weight files distributed via the internet.

DeepSeek's pattern

DeepSeek has now established a consistent pattern: release a model with performance competitive with US frontier systems at significantly lower training cost, with open weights, on Chinese hardware. Each release tests a different assumption behind US AI policy.

V3 tested cost efficiency. V4, if the Huawei chip reports are accurate, tests hardware dependency.

What to watch

The most important data point will be independent benchmark comparisons after V4's release — particularly on the Chatbot Arena (LMSYS), MMLU, and coding benchmarks that researchers use to compare frontier models. If V4 scores competitively with Claude Sonnet, GPT-4o, or Gemini 1.5 Pro, the implications for US export control policy will be immediate and significant.

Also watch for the US government's response. The Biden-era chip restrictions were designed to prevent exactly this scenario. A credible trillion-parameter Chinese model running on Huawei hardware will accelerate debate over whether more aggressive restrictions are needed — or whether restrictions are effective at all.

Source: TechNode