DeepSeek has finally released its much-anticipated next-generation foundational artificial intelligence (AI) model, the open-source V4, which it said was competitive with leading U.S. closed-source models from OpenAI and Google DeepMind. The V4-pro model has 1.6 trillion parameters, making it the company’s biggest-ever model by that metric, while the smaller V4-flash model has 284 billion parameters. A higher parameter count generally correlates with greater capabilities for a model, while also increasing the computational demands of training and serving it. Both models have a context window of 1 million tokens, a critical feature that determines the amount of information an AI system is able to process, which DeepSeek said was achieved with “world-leading” cost efficiency. DeepSeek’s previous flagship model had a context window of 128,000 tokens.
Soon after DeepSeek’s release, Huawei announced “full support” with its Ascend chips, along with its supernode systems, to serve V4 models for model inference. AI chipmaker Cambricon Technologies also moved quickly to announce compatibility with DeepSeek’s new models. “The release of V4 explicitly mentions compatibility with domestic chips,” said analysts from Huatai Securities in a note to clients. “We can look forward to a significant improvement in the capabilities of domestic graphics cards and their widespread adoption this year.” While the parameter size of V4-pro makes it prohibitively large to be run locally on consumer-grade hardware, the extended technical report outlining V4’s model architecture and training techniques is likely to be beneficial for global AI developers.
The V4-flash model is also one of the cheapest cutting-edge models available on the market, with token pricing identical to DeepSeek’s V2 model released in June 2024. The company said that the throughput of V4-pro was currently limited by a shortfall in computational supply, adding that “prices will drop significantly” in the second half of the year “once Huawei’s Ascend 950PR supernodes ship at scale”. Before V4’s release, U.S. officials accused DeepSeek of using banned Blackwell chips from industry leader Nvidia to train its models. DeepSeek did not disclose the hardware stack used to train V4. However, it mentioned the development of “kernels” – codes dictating the functions of graphics processing units (GPUs) – adapted to both Nvidia and Huawei chips in the technical report.
The broader chip supply chain is also set to benefit, with Hangzhou-based securities firm Eastern Communications saying in a research note to clients that the new model is expected to both drive up demand for domestic computing power and accelerate adaptation of domestic chips for leading models, including central processing units (CPUs), the South China Morning Post reports.
Meanwhile, Huawei Technologies has launched its first artificial intelligence glasses, as it joins an intensifying battle with U.S. leader Meta and domestic peers including Alibaba Group Holding and Rokid in smart eyewear. Priced from CNY2,499, the new eyewear weighed just 35.5 grams and featured various AI functions ranging from voice interaction to payments. Powered by Huawei’s self-developed chip designed for eyewear, the glasses enable users to live stream and make video calls in first-person view. With multimodal AI capabilities, they could also estimate and track food calories, and allow payments by scanning a QR code, among other functions. Huawei also unveiled new handsets under its mid-range flagship Pura 90 series, starting from CNY4,699, the same as its predecessor launched last June, but Yu Chengdong, Chairman of Huawei’s consumer business group, said memory shortages had driven up costs by as much as CNY1,500. “We are under a lot of pressure for this pricing strategy,” Yu said. “We may need to raise prices in future when we can’t contain rising costs any longer,” he said. Huawei is one of the few domestic smartphone players that has not yet resorted to price increases, while Oppo, Vivo and Xiaomi have all announced adjustments to absorb rising memory costs.
