August 10, 11:15

Chinese Large Language Models Lead Global Token Usage for Fifteenth Consecutive Week; DeepSeek-V4-Flash Takes First Place

Chinese artificial intelligence laboratories now occupy the top four positions in the global model token usage ranking. According to the latest data from OpenRouter, the total global model usage reached 69 trillion tokens between August 3 and 9, which is 21.48 percent higher compared to the previous week. Chinese models accounted for 34.25 trillion of this total, while American models contributed 9.17 trillion. This marks the fifteenth consecutive week that Chinese models have led the global ranking.

A more detailed analysis of the chart reveals a more complex picture. The official version of DeepSeek-V4-Flash-0731 took first place with 8.83 trillion tokens, representing a 570 percent increase from the previous week. Tencent Hy3 was second with 8.05 trillion tokens, showing a 67 percent growth. The preview version of DeepSeek-V4-Flash-0423 dropped to third place with 5.88 trillion tokens, and Xiaomi MiMo-V2.5 secured fourth place with 5.39 trillion tokens. Thus, all four leading global spots are now held by Chinese labs.

The official release of DeepSeek was made on July 31. The model structure and parameter count remained unchanged compared to the preliminary version, but the fine-tuning recipe was started from scratch. This resulted in significant improvements in mathematical and coding test scores, explaining the rapid shift in developer traffic. According to Artificial Analysis, the official V4-Flash now passes single-task tests at an average cost of $0.03, which is about one hundred times cheaper than Anthropic's flagship Claude Fable 5, making it the most affordable large model globally.

There is also a shift in rankings. The MiniMax M3 model, which was in seventh place last week, and Stepfun Step 3.7 Flash, which was in ninth, have left the top ten in the current ranking. OpenAI's GPT-5.6 Luna entered the top five with 4.43 trillion tokens, showing a 128 percent increase after the company reduced the price for enterprise workloads by 80 percent. Google's Gemini 3.6 Flash entered the top ten for the first time with 2.33 trillion tokens, increasing by 446 percent compared to the previous week.

The structural conclusion is that Chinese open-source models have transitioned from a price advantage to a usage advantage. Fifteen weeks of sustained global leadership, holding four out of the top four spots, and a flagship model priced at one-hundredth—all point to the same conclusion: the Chinese open-source stack no longer competes solely on price. It competes in a combination of price, throughput, and benchmark quality, and wins on all three parameters.

pandaily.comAugust 6, 05:18

Chinese Large Language Models Become Globally Competitive Thanks to DeepSeek Achievements

Initially starting as fan art, a meme featuring DeepSeek founder Liang Wenfen in the guise of a 'cool guy' in a blue suit quickly gained significant resonance on Chinese social media. This meme served as a catalyst for the release of DeepSeek V4 Flash.

The DeepSeek V4 Flash model achieved a score of 7.1 trillion tokens in the weekly global OpenRouter ranking and instantly became the most used model globally. As a result, DeepSeek secured two spots in the world's top five, with four positions held by Chinese models, while GPT-5.6 Luna dropped to sixth place.

Economic indicators show a significant advantage. One day before the release of V4 Flash, OpenAI reduced the price of GPT-5.6 Luna by 80 percent, setting the cost at $0.2 per million input tokens and $1.2 per million output tokens. DeepSeek V4 Flash was released at a price equivalent to approximately $0.14 for input and $0.28 for output, while offering a 98 percent caching discount compared to the industry standard of 90 percent. The independent evaluator Artificial Analysis determined the cost of V4 Flash for a typical agent task to be about 60 percent lower than that of Luna. OpenCode recorded a caching rate of 96 percent, which lowered the average session cost to $0.06.

The performance gap is no longer substantial. DeepSeek's internal tests show that V4 Flash surpasses V4 Pro Preview and GLM 5.2, only trailing Opus-4.8. Furthermore, Artificial Analysis assigned it an intelligence index score of 50, whereas Luna received 51. On the VulcanBench benchmark, Bold Metrics CTO Morgan Linton characterized V4 Flash as 'exceptionally stable under moderate load,' noting its superiority over Claude Fable and GPT-4.5. Approximately half of the traffic directed to DeepSeek on OpenRouter now comes from developers in the US and Europe, indicating that this phenomenon extends beyond the domestic market.

The focus on agent development is deliberate. V4 Flash utilizes the V4 Flash Preview architecture and focuses on agent capabilities after training. Team member Cui Tianyi publicly hired engineers for a new group, Harness—a system that wraps the model with context, tools, task state, and feedback, using open GitHub projects as a selection filter. In May, the DeepSeek-TUI agent, developed by independent American developer Hunter Boone, gathered over 11 thousand stars on GitHub in just a few days. Boone's open letter, titled 'whale bros, I'm the American who builds DeepSeek-TUI,' became a cultural event that swept across the Pacific. Since international developers support these models through their choices, China's open LLMs are no longer just a curiosity; they have become the baseline price-to-performance ratio that all others must now surpass.

Chinese Large Language Models Lead Global Token Usage for Fifteenth Consecutive Week; DeepSeek-V4-Flash Takes First Place

Similar stories

Popular

Chinese Large Language Models Become Globally Competitive Thanks to DeepSeek Achievements