Chinese AI lab DeepSeek has released the official version of its V4-Flash model, the latest move in an intensifying price and performance race among open-source AI developers. The launch, which arrived in early August, follows a preview of the broader V4 family that DeepSeek put out in April.

The V4 lineup ships as two models. V4-Pro is the larger of the pair, built on a Mixture of Experts architecture with 1.6 trillion total parameters and 49 billion active parameters, making it one of the largest open-weight models available anywhere. V4-Flash is the leaner counterpart, running 284 billion total parameters with 13 billion active, and both models offer a 1-million-token context window by default.

On the independent Artificial Analysis Intelligence Index, V4-Pro scored 52, trailing GPT-5.5 at 60 and Claude Opus 4.7 at 57. DeepSeek’s own technical documentation acknowledges the gap directly, stating that V4’s reasoning and agentic capabilities are comparable to models such as GPT-5.2, Gemini 3.0 Pro and Claude Opus 4.5, which were released roughly six months earlier, and that V4 trails the current frontier by about three to six months.

Despite the benchmark gap, analysts see the release as strategically significant. Counterpoint Research principal AI analyst Wei Sun said V4’s profile points to excellent agent capability at a significantly lower cost, a combination that has defined DeepSeek’s approach since its original R1 model rattled markets last year. V4-Flash was released with enhanced agentic features and API pricing up to 50 percent cheaper than earlier versions.

The release lands amid what industry watchers describe as an accelerating price war among Chinese AI labs, with DeepSeek and rivals such as Alibaba’s Qwen competing aggressively on cost even as they close the performance gap with US frontier labs. The Council on Foreign Relations has described DeepSeek V4 as signaling a new phase in the broader US-China AI rivalry, one increasingly shaped as much by pricing strategy as by raw model capability.