Chinese AI company DeepSeek released preview versions of its long-awaited V4 model on April 24th, marking its most significant release since the R1 model upended global AI markets more than a year ago. The company launched two variants - V4 Flash and V4 Pro - both built as mixture-of-experts architectures and both supporting one million token context windows. That context window is large enough to feed an entire codebase or a full set of business documents into a single prompt.
The V4 Pro model carries 1.6 trillion total parameters with 49 billion active per task, making it the largest open-weight model currently available. The mixture-of-experts design means only a fraction of those parameters activate for any given query, keeping inference costs dramatically lower than models of comparable capability. V4 Flash, the smaller variant, carries 284 billion total parameters with 13 billion active.
The Pricing Is the Story
DeepSeek V4 Flash costs $0.14 per million input tokens and $0.28 per million output tokens. V4 Pro comes in at $0.145 input and $3.48 output. Both models undercut GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro on pricing, with analysts at VentureBeat estimating API costs at roughly one-sixth of the leading closed-source alternatives.
For businesses currently paying top-dollar for frontier model access, this matters immediately. Enterprise AI spending decisions that were cost-prohibitive six months ago become viable when the underlying model costs a fraction of what it did. The question is no longer purely performance - it is performance per dollar at scale.
On benchmark tests, DeepSeek says V4 Pro Max outperforms its open-source peers across reasoning categories and surpasses GPT-5.2 and Gemini 3.0 Pro on some tasks. The models trail slightly on pure knowledge benchmarks and fall behind the latest GPT-5.4 and Gemini 3.1 Pro on some zero-shot reasoning tests, suggesting a development trajectory that analysts estimate lags frontier models by roughly three to six months.
The Geopolitical Dimension
The more consequential story here is not the benchmark scores. Huawei confirmed that its Ascend AI processors can support DeepSeek V4. Bloomberg reported that DeepSeek spent months reworking its software stack to optimize specifically for Huawei's Ascend 950PR chips, going beyond basic compatibility into hardware-specific performance tuning.
DeepSeek also gave Chinese chipmakers early optimization access to V4 while withholding that same window from Western silicon suppliers, including Nvidia. This is a deliberate strategic choice, not a technical constraint. If China's leading AI model is optimized first for domestic chips, it accelerates the development of a parallel AI infrastructure that does not depend on Nvidia or American compute.
The launch arrived one day after the US government formally accused China of stealing American AI lab intellectual property at industrial scale using proxy accounts - including allegations against DeepSeek specifically from both Anthropic and OpenAI. The timing was not subtle. DeepSeek V4 is both a technical release and a geopolitical statement.
For business leaders evaluating open-source AI options, DeepSeek V4 is worth taking seriously on cost and capability grounds. The governance, provenance, and security considerations require careful review before deploying on sensitive workloads. But the cheap-good-fast triangle in AI just shifted again.
