Chinese AI startup DeepSeek released preview versions of its new flagship open source model, V4, on Friday. The company says the new model brings improved reasoning, lower costs, and a massive context window.
DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length.
DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models.
— DeepSeek (@deepseek_ai) April 24, 2026
DeepSeek released two versions: V4-Pro and V4-Flash. The Pro version carries 1.6 trillion parameters. The Flash version is a lighter model with 284 billion parameters, designed to be more efficient and affordable.
Both models support a context window of one million tokens. That means they can process a large amount of text at once, putting them on par with Google’s Gemini in that area.
DeepSeek said the models are currently text-only. The company added it is working on adding multimodal capabilities, which would allow the models to process images and video in the future.
On MMLU-Pro, a widely used AI benchmark, V4-Pro matched OpenAI’s GPT-5.4. It came in slightly behind Google’s Gemini and Anthropic’s Claude Opus 4.6. In reasoning benchmarks, V4-Pro trails only the latest Gemini model.
DeepSeek also said V4 has been optimized for AI agent tools including Claude Code, OpenCode, and CodeBuddy.
The company described V4’s context length as “world leading with drastically reduced compute and memory costs.” Analyst Zhang Yi called it an “inflection point,” saying ultra-long context support could move from research labs into everyday commercial use.
AI analyst Max Liu said the release was a “milestone” for China’s AI industry, comparing its potential impact to when DeepSeek’s R1 first launched.
This is DeepSeek’s first major ground-up model release since R1 launched in early 2025. That model shook global tech stocks, including Nvidia and Meta, by showing that a cheaper, leaner model could compete with expensive closed-source rivals.
DeepSeek did not say which chips were used to train V4. Earlier this year, U.S. officials accused the company of using banned Nvidia Blackwell chips. A report from The Information said the models were trained on Huawei chips instead.
Huawei confirmed its Ascend supernode, running Ascend 950 AI chips, would fully support DeepSeek’s V4 models.
The release comes days after reports that Tencent and Alibaba are in talks to invest in DeepSeek at a valuation above $20 billion. DeepSeek is considered one of China’s six leading AI unicorns.
A preview version of V4 is now available on Hugging Face. DeepSeek has not announced a date for the full release.
The post DeepSeek Is Back — And Its New V4 Model Could Shake Up AI Stocks Again appeared first on CoinCentral.
