Chinese artificial intelligence startup DeepSeek has officially released its V4-Flash model, moving the system out of preview status on July 31, 2026, in a launch that underscores the accelerating pace of competition within China’s booming AI sector.
The model, a 284-billion-parameter system with 13 billion active parameters, retains the same underlying architecture unveiled in an April preview but benefits from an extensively refined post-training process, according to the company. Notably, DeepSeek chose to push the smaller Flash variant to official release ahead of its flagship V4-Pro model, a 1.6-trillion-parameter system that remains in preview, suggesting the company is prioritising efficiency and cost-effective deployment over raw scale for now.
On performance, DeepSeek reported strong scores on agent-focused benchmarks, including a result of 82.7 on Terminal-Bench 2.1 and 54.4 on DeepSWE, figures that remain vendor-generated and have not yet been independently verified. Independent research firm Artificial Analysis assigned the model an Intelligence Index score of 50 on a broader measure of general capability. While V4-Flash trails Moonshot AI’s Kimi K3, a 2.8-trillion-parameter rival model, it had already established itself as the most-used model on the OpenRouter platform for seven consecutive weeks earlier this summer, reflecting strong developer adoption even before the official release.
Pricing is central to DeepSeek’s strategy. The company cut API costs by as much as 50%, setting rates at $0.0028 per million cache-hit input tokens, $0.14 per million cache-miss input tokens, and $0.28 per million output tokens, roughly a third of what it charges for the flagship V4-Pro model. The model supports a one-million-token context window with up to 384,000 tokens of output, and can handle 2,500 concurrent requests, five times the capacity of the Pro version, a significant advantage for developers running parallel AI agent workloads. A previously announced dynamic pricing mechanism that would double costs during peak hours has still not been implemented.
The release lands alongside major financial news for the company. DeepSeek has closed a funding round of nearly $7.4 billion, valuing the startup at 350 billion yuan (roughly $49 billion), with backing from technology giants Tencent and NetEase. The company has said it plans to double its headcount and sharpen its focus on developing autonomous AI agents, a segment increasingly seen as the next competitive battleground among Chinese AI developers.
DeepSeek’s move comes as Chinese AI firms race to close the gap with, and in some benchmarks surpass, their American counterparts, while undercutting them sharply on price. The company’s earlier V3 and R1 models triggered significant disruption in global AI markets in early 2025, prompting renewed scrutiny of the balance of power between Silicon Valley and Chinese AI labs. Rivals such as Alibaba, Moonshot AI, and Zhipu AI have each unveiled competing large language models in recent months, intensifying a domestic race that is increasingly spilling into international markets through low-cost API offerings.
As of the official V4-Flash release, no corresponding model weights had appeared on Hugging Face, leaving open-source availability of the new version uncertain for now, a detail that developers and researchers tracking the open-weight AI ecosystem will be watching closely in the coming weeks.
