DeepSeek Previews V4, Its Biggest Open-Weight Model Yet, at Prices That Undercut the Field
The Chinese AI lab's V4 Pro model tops 1.6 trillion parameters, claims near-parity with frontier models on reasoning, and costs a fraction of GPT-5.5 or Claude Opus.
Chinese AI lab DeepSeek on Friday released preview versions of its newest large language model, DeepSeek V4, in what the company says nearly closes the performance gap with leading closed-source models. The release consists of two variants: V4 Flash, a lightweight model optimized for speed, and V4 Pro, which at 1.6 trillion total parameters is the largest open-weight model publicly available.
Both models are mixture-of-experts architectures that activate only a subset of parameters per task, dramatically reducing inference costs. They support context windows of one million tokens, large enough to process entire codebases or lengthy legal documents.
DeepSeek says the models are more efficient and performant than its prior V3.2 release due to architectural improvements, and that V4-Pro-Max outperforms open-source peers across reasoning benchmarks and surpasses OpenAI's GPT-5.2 and Gemini 3.0 Pro on some tasks.
On pricing, V4 Flash is listed at $0.14 per million input tokens and $0.28 per million output tokens, undercutting GPT-5.4 Nano, Gemini 3.1 Flash, and Claude Haiku 4.5. V4 Pro comes in at $0.145 per million input and $3.48 per million output, below Gemini 3.1 Pro, GPT-5.5, and Claude Opus 4.7.
However, both models lag frontier models on knowledge-intensive benchmarks, suggesting a developmental trajectory trailing state-of-the-art by roughly three to six months.
The release arrives against a backdrop of heightened geopolitical tension. OpenAI and Anthropic have previously accused DeepSeek of improperly using their model outputs through a process known as distillation, and the U.S. State Department issued a global diplomatic warning on the same day alleging that DeepSeek and other Chinese AI firms engaged in industrial-scale theft of American AI intellectual property.
DeepSeek did not immediately respond to requests for comment.
Read the original reporting at TechCrunch.
