On April 24, DeepSeek, the Chinese AI startup company, launched the first preview version of its long-awaited V4 models running entirely on Huawei Ascend chips. The new tech creations debuted as a cost-efficient rival to OpenAI and Google Gemini.
DeepSeek V4 model series has been in rumors for quite some time. It was expected to see a launch delay to next month. But the company has finally unveiled them.
The open-source DeepSeek V4 series has two models: V4-Pro with around 1.6 trillion parameters and the V4-Flash with 284 billion parameters. Here, the higher parameter count suggests greater AI capabilities with increased computational demands.
AI V4 Series Models:
- DeepSeek V4-Flash: 284 billion parameters + 13 billion active. It is a faster, more cost-efficient version for rapid inference.
- DeepSeek V4-Pro: 1.6 trillion, 49 billion. It is a flagship and is more optimized for complex reasoning, math, and coding.
Note that the preview version refers to an early-stage release. It’s basically a first look before a full and stable launch. The models are open for testing and integration publicly, but aren’t the final products.
V4-Pro with 1.6 trillion parameters is the company’s biggest model ever by metrics. Both DeepSeek V4 models are running on the Huawei Ascend AI chips, underscoring China’s “resilient growth” in the AI tech industry.
(Image Credits: DeepSeek/X)
Both DeepSeek V4 models come with 1 million tokens. Here, token refers to the fundamental units of data – words, characters, or subwords that AI models process to understand and generate language.
For comparison, the previous DeepSeek model has a context window of 128,000 tokens. But the company says that the latest version has achieved world-leading cost-efficiency. In a note to clients, Huatai Securities analysts said:
“The release of V4 explicitly mentions compatibility with domestic chips. We can look forward to a significant improvement in the capabilities of domestic graphics cards and their widespread adoption this year.”
Notably, the DeepSeek V4 Pro is quite large to run locally on consumer-level devices. However, its architecture and training techniques are likely to be beneficial for global AI developers.
The company hasn’t mentioned the exact specs and hardware components of the V4 AI models. Though it said that the “kernels” development codes dictate the functions of the GPU adapted to both Nvidia and Huawei chips.
DeepSeek stated:
“We will always adhere to the principle of long-termism, move forward steadily through trial and error and reflection, and strive to get closer to the goal of achieving artificial general intelligence.”
The post DeepSeek launches new V4 AI models running on Huawei chips appeared first on Huawei Central.
