DeepSeek releases V4 AI model; Huawei pledges chip support

On Friday, DeepSeek, an AI startup located in Hangzhou, unveiled its advanced open-source foundational model V4, asserting its competitiveness with leading closed-source models from OpenAI and Google DeepMind. The launch included two variants: V4-pro, boasting 1.6 trillion parameters, and V4-flash, which contains 284 billion parameters. Both models are equipped with a context window of 1 million tokens, a significant upgrade from the 128,000-token window of the previous flagship model, achieved through what DeepSeek calls "world-leading" cost efficiency. Huawei promptly expressed "full support" for its Ascend chips and supernode systems for V4 model inference, with additional details to be disclosed in a livestream later that day. Cambricon Technologies also confirmed compatibility. Analysts at Huatai Securities highlighted V4's explicit mention of compatibility with domestic chips, predicting a notable enhancement in domestic graphics card performance and widespread adoption this year. While the V4-pro is too large for consumer hardware, its technical report is expected to aid global AI developers. The V4-flash model is priced competitively, matching the token pricing of DeepSeek's V2 model from June 2024.