Chinese artificial intelligence company DeepSeek on Friday released preview versions of its newest flagship models, marking its most significant ground-up release since the R1 model rattled global markets in early 2025. The company introduced two variants — DeepSeek V4-Pro and DeepSeek V4-Flash — posting details on the open-source platform Hugging Face.
Both releases highlight advances in reasoning, coding performance, and processing capacity.
Two models with distinct goals
The DeepSeek V4-Pro model carries 1.6 trillion parameters, placing it among the largest open-source AI systems in existence. The V4-Flash version scales down to 284 billion parameters, targeting faster response times with lower computing demands.
Both models share a context window of 1 million tokens — the measurement that determines how much information a model can process in a single exchange. That figure dwarfs most available systems.
Text processing drives both models for now. DeepSeek confirmed it was “working on incorporating multimodal capabilities,” a development that would expand the models to handle images and video.
Performance narrows the gap
DeepSeek made bold claims about its most advanced variant.
The company said Deepseek V4-Pro-Max achieved “top-tier performance in coding benchmarks and significantly bridges the gap with leading closed-source models on reasoning and agentic tasks.”
Benchmark data backs that up, at least partially. On MMLU-Pro — a widely used evaluation for reasoning and coding ability — DeepSeek V4-Pro matched OpenAI’s GPT-5.4. It trailed slightly behind Google’s Gemini-3.1-Pro and Anthropic’s Claude Opus 4.6, according to DeepSeek’s own figures.
That performance matters. Open-source systems have consistently struggled to keep pace with proprietary models. DeepSeek’s latest results suggest that the gap is narrowing faster than many expected.
Hardware sourcing under scrutiny
DeepSeek did not disclose which processors it used to train the V4 models. That omission draws attention.
Earlier in 2026, U.S. officials alleged the company used banned NVIDIA Blackwell chips during development. DeepSeek has not confirmed or denied those claims.
Recent reporting from The Information pointed elsewhere, suggesting the company trained its latest models on advanced chips from Huawei. Separately, Huawei announced that its Ascend supernode — built around the company’s flagship Ascend 950 AI chips — would fully support the V4 lineup.
If accurate, that partnership signals a broader strategic shift. China’s AI sector appears to be accelerating its move away from U.S. semiconductor technology, building a more self-contained development pipeline.
R1 set the stage
Friday’s release arrives roughly 15 months after DeepSeek’s R1 model upended the industry. R1 delivered performance comparable to top closed-source systems while using fewer resources — a combination that shocked markets and triggered sharp declines in major technology stocks.
Investors questioned whether enormous AI infrastructure spending still made sense when a leaner model could produce similar results. The selloff that followed underscored how deeply DeepSeek had disrupted conventional thinking about AI development costs.
V4 builds on that foundation. It reinforces the argument that open-source development can remain competitive without the resource overhead of proprietary systems.
Investor interest climbs
The V4 launch arrives days after reports that Tencent and Alibaba are in active talks to invest in DeepSeek. Those discussions reportedly value the company at more than $20 billion.
That figure positions DeepSeek among the most valuable AI startups in the world. It also reflects growing confidence — both domestically and internationally — in the company’s long-term potential.
DeepSeek already holds a prominent place among China’s so-called “AI Tigers,” a group of six AI unicorns leading the country’s ambitions in the sector. The DeepSeek V4 release reinforces that standing.
What does open-source access change?
DeepSeek’s commitment to open-source development sets it apart from competitors that keep their models locked behind proprietary systems. Open access lets developers adapt, extend, and build on the technology without licensing barriers.
That approach accelerates adoption. It also compresses the timeline between research breakthroughs and real-world applications — in software, enterprise automation, and beyond.
The competitive implications run deep. If open-source models continue to match closed ones on performance, pricing structures, and business models across the industry face serious pressure.
What comes next
DeepSeek’s immediate roadmap centers on multimodal expansion. Adding image and video processing would move the V4 models closer to full-spectrum AI capability — a threshold that currently separates specialized tools from general-purpose systems.
Regulatory scrutiny also looms. Questions around chip sourcing, compliance with export controls, and international competitiveness will shape how the company operates going forward.
Still, the trajectory points in one direction. DeepSeek is moving fast, and the distance between open and closed AI continues to shrink.
Does DeepSeek’s V4 shift the balance of power in artificial intelligence — or does the hardware question change the picture? Please drop your views below.