Key Highlights
- DeepSeek unveiled two open source AI models: V4-Pro (1.6T parameters) and V4-Flash (284B parameters)
- Each model features a 1 million token context window, matching capabilities found in Google’s Gemini
- V4-Pro achieves performance equivalent to OpenAI’s GPT-5.4 on coding benchmarks while ranking second only to Gemini for reasoning tasks
- The company emphasizes significantly reduced compute and memory costs when compared to competing solutions
- This launch coincides with reports of Tencent and Alibaba discussing potential investment in DeepSeek at valuations exceeding $20B
Chinese artificial intelligence startup DeepSeek introduced preview editions of V4, its latest flagship open source model, this past Friday. The company highlights enhanced reasoning capabilities, reduced operational expenses, and extensive context processing as key features of this release.
The company launched two distinct versions: V4-Pro and V4-Flash. The Pro variant operates with 1.6 trillion parameters, while the Flash version represents a streamlined alternative featuring 284 billion parameters, engineered for enhanced efficiency and cost-effectiveness.
Each model handles context windows spanning one million tokens. This capability enables them to analyze substantial volumes of text simultaneously, positioning them alongside Google’s Gemini in processing capacity.
DeepSeek confirmed the models currently function as text-only systems. The organization stated that development efforts continue toward incorporating multimodal capabilities, which will enable the models to analyze images and video content down the line.
Performance Against Major Competitors
When evaluated using MMLU-Pro, a commonly referenced AI assessment tool, V4-Pro delivered results equal to OpenAI’s GPT-5.4. Performance fell marginally below Google’s Gemini and Anthropic’s Claude Opus 4.6. For reasoning-specific benchmarks, V4-Pro secured the second-highest position, trailing only the most recent Gemini release.
DeepSeek highlighted that V4 has received optimization for AI agent platforms such as Claude Code, OpenCode, and CodeBuddy.
The organization characterized V4’s context length as “world leading with drastically reduced compute and memory costs.” Industry analyst Zhang Yi identified this development as an “inflection point,” suggesting ultra-long context capabilities may transition from experimental settings into mainstream commercial applications.
AI industry analyst Max Liu described the launch as a “milestone” for China’s artificial intelligence sector, drawing parallels to the industry impact generated when DeepSeek’s R1 initially debuted.
Investment Landscape and Market Position
This represents DeepSeek’s first comprehensive model built from scratch since R1 appeared in early 2025. That previous release sent ripples through global technology markets, affecting companies including Nvidia and Meta, by demonstrating how a more economical, efficient model could challenge expensive proprietary alternatives.
DeepSeek declined to specify which processing chips powered V4’s training. Earlier this year, U.S. authorities alleged the company utilized prohibited Nvidia Blackwell chips. A report from The Information indicated the models underwent training using Huawei chips as an alternative.
Huawei verified that its Ascend supernode, powered by Ascend 950 AI processors, would provide comprehensive support for DeepSeek’s V4 models.
The model’s arrival follows recent reports indicating Tencent and Alibaba have entered discussions regarding investment in DeepSeek at valuations surpassing $20 billion. DeepSeek holds recognition as one of China’s six premier AI unicorn companies.
A preview build of V4 became accessible through Hugging Face. DeepSeek has yet to disclose a timeline for the complete public release.
