Key Highlights
- DeepSeek unveiled two open-source AI models: V4-Pro featuring 1.6 trillion parameters and V4-Flash with 284 billion parameters
- Each model offers one million token context capacity, matching Google Gemini’s capabilities
- V4-Pro achieves performance parity with OpenAI’s GPT-5.4 in coding tests and ranks second only to Gemini in reasoning tasks
- The company emphasizes significantly lower computational and memory requirements versus competitors
- Launch follows news of potential $20B+ investment talks with Tencent and Alibaba
On Friday, Chinese artificial intelligence company DeepSeek introduced preview editions of V4, its latest flagship open-source model. According to the startup, this new offering delivers enhanced reasoning capabilities, reduced operational expenses, and an exceptionally large context window.
The company launched two distinct variants: V4-Pro and V4-Flash. The Pro edition contains 1.6 trillion parameters, while Flash represents a streamlined alternative featuring 284 billion parameters, optimized for efficiency and cost-effectiveness.
Each variant supports processing up to one million tokens simultaneously. This capability enables them to handle substantial volumes of text in a single operation, positioning them alongside Google’s Gemini in this particular specification.
DeepSeek noted that both models currently function with text exclusively. The organization stated it’s actively developing multimodal features that will enable image and video processing down the line.
Performance Against Competitors
In MMLU-Pro testing, a standard industry benchmark, V4-Pro delivered results equivalent to OpenAI’s GPT-5.4. It performed marginally below Google’s Gemini and Anthropic’s Claude Opus 4.6. For reasoning assessments, V4-Pro secured second place, trailing only the most recent Gemini release.
DeepSeek further indicated that V4 has received optimization for AI agent platforms such as Claude Code, OpenCode, and CodeBuddy.
The organization characterized V4’s context capacity as “world leading with drastically reduced compute and memory costs.” Industry analyst Zhang Yi labeled it an “inflection point,” suggesting that extended context capabilities might transition from experimental settings into mainstream commercial applications.
AI specialist Max Liu characterized the launch as a “milestone” for China’s artificial intelligence sector, drawing parallels to the impact of DeepSeek’s original R1 debut.
Financial and Industry Landscape
This marks DeepSeek’s first significant new-generation model introduction since R1 debuted in early 2025. That previous release disrupted global technology markets, affecting companies like Nvidia and Meta, by demonstrating that an economical, streamlined model could rival costly proprietary alternatives.
DeepSeek hasn’t disclosed which processing units powered V4’s training. Earlier in the year, American authorities alleged the company utilized prohibited Nvidia Blackwell processors. A subsequent report from The Information indicated the models were instead trained using Huawei hardware.
Huawei verified that its Ascend supernode platform, equipped with Ascend 950 AI processors, would provide complete compatibility for DeepSeek’s V4 models.
The announcement arrives shortly after media reports revealed that Tencent and Alibaba are conducting investment discussions with DeepSeek at a valuation exceeding $20 billion. The company ranks among China’s six premier AI unicorns.
An early-access version of V4 is currently accessible through Hugging Face. DeepSeek hasn’t revealed a timeline for the complete public rollout.
