Key Takeaways
- DeepSeek unveiled two open-source AI models on Friday: V4-Pro featuring 1.6 trillion parameters and V4-Flash with 284 billion parameters
- Each model offers a context window capable of handling 1 million tokens, matching capabilities seen in Google’s Gemini
- On coding performance tests, V4-Pro performs comparably to OpenAI’s GPT-5.4 and ranks second only to Gemini in reasoning tasks
- The company highlights significant reductions in computational and memory requirements versus competitor models
- This launch coincides with reports that Tencent and Alibaba are negotiating investments in DeepSeek at valuations exceeding $20 billion
On Friday, Chinese artificial intelligence company DeepSeek introduced preview builds of its latest flagship open-source model family, designated V4. According to the firm, these new releases deliver enhanced reasoning capabilities, reduced operational expenses, and an expansive context processing capacity.
The company launched two distinct variants: V4-Pro and V4-Flash. The Pro configuration incorporates 1.6 trillion parameters, while the Flash alternative represents a streamlined architecture containing 284 billion parameters, optimized for enhanced efficiency and cost-effectiveness.
Both variants feature support for processing one million tokens in a single context window. This capability enables them to analyze substantial volumes of text simultaneously, positioning them competitively alongside Google’s Gemini offering in this dimension.
DeepSeek noted that the current models exclusively handle text-based inputs. The organization stated it is actively developing multimodal functionality, which will enable future versions to interpret images and video content.
Performance Against Competing Models
When evaluated using MMLU-Pro, a standard AI performance metric, V4-Pro achieved results comparable to OpenAI’s GPT-5.4. It registered marginally lower scores than Google’s Gemini and Anthropic’s Claude Opus 4.6. For reasoning evaluations, V4-Pro secured second place, trailing only the most recent Gemini release.
DeepSeek further indicated that V4 has undergone specific optimization for integration with AI agent frameworks 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 ultra-long context capabilities might transition from experimental research environments into mainstream commercial applications.
AI industry observer Max Liu characterized the launch as a “milestone” for China’s artificial intelligence sector, drawing parallels to the market impact generated by DeepSeek’s initial R1 model release.
Financial and Investment Landscape
This marks DeepSeek’s first comprehensive new model architecture since R1 debuted in early 2025. That previous release created significant turbulence in global technology markets, affecting companies including Nvidia and Meta, by demonstrating that a more economical, efficient model could rival expensive proprietary alternatives.
DeepSeek has not disclosed which semiconductor hardware was utilized for V4 training. Earlier in the year, United States authorities alleged the company employed restricted Nvidia Blackwell processors. Subsequently, The Information published findings indicating the models were actually trained using Huawei-manufactured chips.
Huawei has verified that its Ascend supernode infrastructure, powered by Ascend 950 AI processors, will provide comprehensive support for DeepSeek’s V4 model family.
The model announcement follows closely behind media reports indicating Tencent and Alibaba have entered discussions regarding potential investments in DeepSeek, with proposed valuations surpassing $20 billion. Industry observers categorize DeepSeek among China’s six premier AI unicorn companies.
An early-access preview build of V4 is currently accessible through Hugging Face. DeepSeek has yet to confirm a timeline for the official production release.


