The official version of DeepSeek-V4-Flash, a lightweight model with 304 billion parameters, was released by DeepSeek on August 2. This version demonstrates significant superiority over the preliminary V4-Pro after minimal fine-tuning and achieves performance levels comparable to Claude Opus-4.8.
The model weights were opened to the community and immediately ranked second in HuggingFace trends, only behind Kimi K3. According to community benchmarks, the model showed a performance score of 82.7 at a cost of only $0.14 per million input tokens and $0.28 per million output tokens, offering a price-to-quality ratio significantly better than leading closed models.
Independent testing on the VulcanBench platform placed DeepSeek-V4-Flash first, surpassing Claude Fable 5. This release has been widely recognized in the industry as 'DeepSeek's third moment,' following DeepSeek-R1, which triggered the first global AI price review in January 2025, and the K3 release cycle, which changed the landscape of open weights.
The official V4-Flash version was analyzed using the new DeepSeek Harness structure—a toolkit designed for structured evaluation of agent software engineering tasks. The Harness framework provides standardized interfaces for verifying tool contract adherence, linking evidence to conclusions, and verification workflows, indicating DeepSeek's expansion beyond simple model releases into AI coding infrastructure.
The testing methodology used to validate V4-Flash's performance was also based on DeepSeek Harness, suggesting the company has developed an agent orchestration and evaluation layer that could form the basis of a future DeepSeek Code product. The AI coding market could undergo fundamental changes if DeepSeek combines V4-class model capabilities with an open evaluation framework and aggressive pricing.
Furthermore, the release confirmed that the official DeepSeek-V4-Pro version is under development. Given that the lightweight Flash variant is already approaching the performance of advanced closed systems, it is expected that the full Pro model will significantly boost capability metrics. DeepSeek has not yet announced a release timeline for the official V4-Pro version. The company continues its strategy of opening weights while maintaining competitive pricing, positioning DeepSeek as a benchmark for the Chinese open-source AI ecosystem and a price anchor in the global API market.
This third moment from DeepSeek reflects the accelerating pace of Chinese open-weight model releases. Kimi K3 with 2.8 trillion parameters sets the scale ceiling for open source, MiniMax H3 introduced full multimodal generation at one-third the cost of competitors, and now DeepSeek-V4-Flash demonstrates that a 304-billion-parameter model, trained rigorously, can approach state-of-the-art performance. The convergence of open weights, Harness tools, and aggressive pricing increases competitive pressure on closed solution providers and alters developers' approaches to model selection. The official DeepSeek-V4-Flash with open weights and the Harness framework may signal DeepSeek's transition from a research lab to a full-fledged player in AI infrastructure, with AI coding positioned as the next battleground where model capabilities, evaluation infrastructure, and developer user experience converge.
CFMOTO has reached a significant milestone in the development of its future superbikes. The V4 SR-RR prototype demonstrated a speed of 315.82 km/h during an official test.
The test took place as part of the China Motorcycle Top Speed Challenge on June 15th at the CCCC Automotive Test Center in Shangrao, China. The speed, officially measured by GPS, made the V4 SR-RR the fastest internal combustion engine motorcycle in China in history.
The V4 SR-RR, which was presented as a concept at EICMA 2025, is equipped with a new 90-degree V4 engine with a displacement of 997 cm³. This engine is capable of developing around 210 horsepower. According to CFMOTO data, the project also features a power-to-weight ratio of less than 1 kg per horsepower, a characteristic usually found in the most extreme models in this category.
The performance metrics of this Chinese sport model put it on par with well-known models such as the Ducati Panigale V4 S with 213 hp and the Aprilia RSV4 Factory with 216 hp. Although the model remains a prototype, the similar power achieved with a smaller engine displacement draws attention to the manufacturer's technical progress.
In addition to absolute figures, the comparison with racing is impressive. The speed of 315.82 km/h lags only slightly more than 50 km/h behind the recent MotoGP record of 368.6 km/h set in Mugello.
The V4 SR-RR is still under development, and official release dates are not yet available. It is expected that this motorcycle will serve as the basis for CFMOTO's future participation in the World Superbike Championship (WorldSBK), strengthening the brand's position in the high-performance motorcycle segment.
