- New model delivers major gains in coding and agent benchmarks at a fraction of rival costs
- Company reportedly seeks $71 billion valuation while preparing for a potential STAR Market listing
DeepSeek on July 31 launched the public beta of its official DeepSeek-V4-Flash API, marking the long-awaited release of the production version of its latest AI model.
The final version keeps the same architecture as the preview release but achieves a significant leap in coding and agent capabilities through additional post-training, according to the company.
Benchmark results showed sharp improvements. On Artificial Analysis’ intelligence index, V4-Flash-0731 scored 50 points, six points higher than the V4-Pro preview version.
In the Frontend Code Arena, the model reached 1,586 points to top the leaderboard, improving by 154 points from the preview version.
Its performance on the DeepSWE benchmark showed the biggest jump, rising from 7.3 points in the preview version to 54.4 points, an increase of more than sixfold.
The model also maintains DeepSeek’s low-cost strategy. Its output pricing stands at around $0.28 per million tokens, roughly one-ninetieth the price of Claude Opus 4.8.
The company also introduced its self-developed Agent testing framework Harness, marking the first official disclosure of capabilities from the newly formed Harness team.
Funding and IPO plans accelerate
DeepSeek is reportedly preparing a new financing round just one month after completing its first fundraising, with a pre-money valuation of around $71 billion, according to multiple media reports.
The company has also hired investment banks to prepare for a potential listing on the Shanghai Stock Exchange’s STAR Market, with plans to submit an application this year, the reports said.
Unlike the previous round, the new financing would allow participation from overseas investors through the Qualified Foreign Limited Partner (QFLP) mechanism, report said.
Earlier in July, DeepSeek had reportedly suspended its second found round due to what media said is founder and CEO Liang Wenfeng’s frustration over the leak of his internal meeting with some shareholders.
The push for fresh capital comes as frontier AI development requires increasingly heavy investment in computing infrastructure.
DeepSeek is reportedly building gigawatt-scale computing facilities, with self-operated data centers potentially requiring investment worth hundreds of billions of yuan.
Breaking the ‘bigger is better’ assumption
The release of V4-Flash once again challenges the industry belief that larger models automatically deliver superior performance.
With only one-sixth of the total parameters and less than one-quarter of the active parameters of its V4-Pro preview version, V4-Flash has surpassed its predecessor on multiple benchmarks.
The approach echoes DeepSeek’s breakthrough in early 2025, when the company demonstrated that high-performance AI models could be trained at significantly lower costs than many global competitors.
Tianfeng Securities, a stock brokerage firm, described the launch as a “major upside surprise,” saying it could benefit China’s domestic AI model ecosystem.
On the same day as V4-Flash’s release, OpenAI reportedly cut the price of its GPT-5.6 Luna model by 80%, a move some market observers viewed as a response to DeepSeek’s aggressive cost-performance strategy.
For the global AI industry, DeepSeek’s latest release highlights a growing shift: competition is moving beyond simply building larger models toward improving efficiency, reducing inference costs and turning AI agents into practical productivity tools.
