Particle.news
Download on the App Store
33 ARTICLES
·
last wk.

DeepSeek Releases Preview of V4, an Open-Source AI Model With a 1 Million-Token Context

The launch tests China’s bid to pair low-cost, long-context AI with homegrown Huawei chips under growing U.S. scrutiny.

FILE - The smartphone apps DeepSeek page is seen on a smartphone screen in Beijing, Jan. 28, 2025. (AP Photo/Andy Wong, File)
A citizen is watching a display board about DepSeek AI big model in a shopping mall in Hangzhou, Zhejiang Province, China on April 23, 2026.
Image
Image

Overview

  • DeepSeek released the V4 preview on Friday, offering V4-Pro with 1.6 trillion parameters and V4-Flash with 284 billion, each with a one million-token context and open-source access.
  • DeepSeek claims V4-Pro tops other open models on math and coding and trails leading closed systems by a small margin, with usage prices that undercut comparable models from OpenAI, Google and Anthropic.
  • Both models focus on long tasks and autonomous agent workflows such as coding, remain text-only for now, and are tuned for tools like Claude Code, OpenClaw and CodeBuddy.
  • Huawei said its Ascend Supernode clusters will support V4, a move that advances China’s push to run major AI models on domestic chips rather than relying on U.S. hardware.
  • U.S. officials have warned of industrial-scale model distillation by Chinese firms and media reports have alleged chip restrictions were skirted, and independent benchmarking and regulatory reviews now loom large for V4’s adoption.