DeepSeek Open-Sources V4 Model Series, Disrupting AI Accessibility

In a significant move for the AI community, DeepSeek launched the preview versions of its V4 model series on April 24th, releasing them as open-source. The release features two primary variants: the high-performance DeepSeek-V4-Pro and the efficient DeepSeek-V4-Flash, catering to different application needs.

Technical Milestone: Fine-Grained Expert Parallelism Goes Cross-Platform

Official technical documentation highlights a major achievement: the successful validation of DeepSeek V4's Fine-Grained Expert Parallelism (EP) scheme. This innovative architecture has been proven to work seamlessly on two major AI hardware platforms: the industry-standard Nvidia GPUs and Huawei's Ascend NPUs. This cross-platform compatibility significantly enhances the model's deployment flexibility and future-proofing.

Current Constraints and a Promising Price Forecast

The company's API information page candidly addresses a current limitation. Due to the constrained availability of high-end computing power, the service throughput for the V4-Pro model is currently limited, potentially affecting large-scale deployment.

The outlook, however, is set to change dramatically. DeepSeek indicates a clear expectation: in the second half of this year, with the mass production and rollout of Huawei's Ascend 950 super nodes, the supply of premium AI compute will see a substantial increase. This infrastructure advancement is predicted to directly trigger a "significant reduction" in the pricing for the V4-Pro model's API services. This development promises to lower the barrier to entry for accessing cutting-edge AI capabilities for developers and businesses alike.

  • Key Event: DeepSeek V4 series preview launched and open-sourced.
  • Tech Breakthrough: Fine-Grained Expert Parallelism validated on Nvidia and Huawei Ascend chips.
  • Present Challenge: Pro model throughput limited by scarce high-end compute.
  • Future Outlook: Ascend 950 mass production expected to slash Pro model pricing.