San Francisco: Chinese AI startup DeepSeek has unveiled preview versions of its newest AI model, saying it matches the performance of the world’s most advanced AI systems while being far cheaper to run.
The new model, called DeepSeek V4, comes in two preview versions, namely V4 Flash and V4 Pro. It is the company’s first major update since its V3.2 and R1 models gained global attention last year for offering strong performance at low cost.
Both versions of V4 are mixture‑of‑experts models, a design that activates only parts of the system for each task. This method keeps operating costs lower while still delivering strong results.
DeepSeek says both models support a context window of one million tokens, allowing users to process very large documents, long conversations, or full software codebases in a single prompt.
The larger V4 Pro model has a total of 1.6 trillion parameters, with 49 billion active at a time. This makes it the largest open‑weight AI model available today, surpassing competitors from other Chinese labs. The smaller V4 Flash model, on the other hand, uses 284 billion parameters and is designed for faster, lighter workloads while still benefiting from the large context window.
According to DeepSeek, both models show major improvements over earlier versions due to architectural changes. The company says V4 has almost “closed the gap” with frontier models such as OpenAI’s GPT series and Google’s Gemini on reasoning tasks.
In coding benchmarks, DeepSeek reports that V4’s performance is comparable to GPT‑5.4. The company also claims its V4 Pro model outperforms other open‑source models on several reasoning tests.
However, DeepSeek acknowledged that its models still lag slightly behind leading systems in general knowledge tests. The company estimates that its development timeline is about 3-6 months behind the most advanced closed‑source models.
Unlike many premium AI systems, DeepSeek V4 supports text only. It does not handle images, audio, or video like some competing frontier models.
One major advantage of V4 is price. DeepSeek says V4 Flash costs just USD 0.14 per million input tokens and USD 0.28 per million output tokens, making it cheaper than lightweight offerings from OpenAI, Google, and Anthropic.
The more powerful V4 Pro model also undercuts high‑end competitors such as GPT‑5.5, Gemini 3.1 Pro, and Claude Opus on pricing.
DeepSeek’s preview launch suggests the company is continuing its strategy of delivering high‑performance AI at lower cost. Its earlier R1 model shocked the global tech industry by matching top U.S. systems using far fewer resources.
Previous
