DeepSeek V4-Flash Tops Global AI Usage Rankings as Company Resumes $8 Billion Funding Round

A Chinese Open-Source Model Just Became the Most-Used AI in the World

In a moment that encapsulates how rapidly the balance of power in artificial intelligence is shifting, DeepSeek’s V4-Flash model processed an astonishing 7.22 trillion tokens on OpenRouter during the week of July 27 to August 2, 2026—enough to claim the number-one spot on the platform’s global weekly usage rankings. OpenRouter, which aggregates API access to models from dozens of providers including OpenAI, Anthropic, Google, and Meta, serves as one of the most reliable real-world barometers of which AI models developers and enterprises are actually building with.

The achievement is more than a vanity metric. It signals that DeepSeek—a startup most Western observers hadn’t heard of eighteen months ago—has managed to outpace every frontier model from Silicon Valley’s largest labs in raw deployment volume. Perhaps even more strikingly, Chinese models claimed nine of the top ten slots on OpenRouter’s weekly ranking, according to reporting from Pandaily, underscoring a structural shift in the global AI ecosystem.

V4-Flash 0731: A Technical Powerhouse Built for Efficiency

The model behind the surge is DeepSeek-V4-Flash-0731, officially released on July 31, 2026. It is a Mixture-of-Experts (MoE) architecture with 284 billion total parameters, but it activates only 13 billion parameters per token during inference. This sparse activation strategy is the key to its remarkable cost-efficiency: the model delivers near-frontier reasoning quality at a fraction of the compute cost of dense models of comparable total size.

The architecture also supports a 1 million-token context window, putting it in the same league as Google’s Gemini models for long-document processing. On the coding front, V4-Flash 0731 reportedly achieves 79% on SWE-bench Verified and 91.6% on LiveCodeBench—benchmarks that measure real-world software engineering and competitive programming ability respectively. These are numbers that would have seemed implausible for a model activating only 13 billion parameters just a year ago.

DeepSeek’s post-training pipeline for the 0731 release focused heavily on agentic capabilities. The model was fine-tuned with extensive tool-use trajectories and multi-step reasoning chains, giving it a significant edge on the kind of agentic workflows—code generation, web browsing, function calling—that dominate real-world API usage. According to MarkTechPost, the upgrade brought “major agentic and coding gains” across nine independent benchmark categories, with the Flash variant now outperforming the larger V4-Pro preview on all nine agentic tasks despite being a deliberately smaller and cheaper model.

Pricing That Reshapes the Market

The economics of V4-Flash are almost aggressively competitive. On OpenRouter, the model is priced at approximately $0.08 per million input tokens and $0.252 per million output tokens. By comparison, leading frontier models from Western labs routinely charge $3 to $15 per million tokens. This order-of-magnitude price advantage is a deliberate strategy: DeepSeek has consistently used aggressive pricing to capture developer mindshare, and the OpenRouter rankings suggest it is working.

The model is released under the MIT license, with open weights available on Hugging Face. This open-source approach has made it a favorite among the self-hosting community, with LocalLLaMA users reporting that the 13B active parameters can run in as little as 10-12 GB of VRAM when quantized to MXFP4 precision—a threshold that puts frontier-class inference within reach of consumer hardware.

From Pause to Restart: The $8 Billion Round

While V4-Flash was conquering the usage charts, DeepSeek was simultaneously navigating a complex funding saga. On August 6, 2026, Bloomberg reported that the company had resumed its second major funding round, seeking to raise close to $8 billion at a valuation near $74 billion. Hong Kong-based Monolith Management is reportedly in talks to participate as a lead investor.

The restart came just twelve days after DeepSeek had abruptly paused the same round on July 25. That suspension was triggered by a controversy surrounding leaked comments from founder Liang Wenfeng about the company’s reliance on Nvidia chips—comments that went viral on Chinese social media and reportedly caused unease among prospective investors and potentially among Chinese regulators. The leak highlighted the geopolitical tightrope that Chinese AI companies walk: dependent on advanced Western GPUs for training while simultaneously positioned as national champions in a US-China technology competition.

DeepSeek’s first external funding round, closed in June 2026, raised $7.4 billion (approximately 50 billion yuan) at a valuation exceeding $50 billion. Liang Wenfeng personally contributed roughly $3 billion to that round and retained total control of the company through an unusual governance structure—a rarity in the venture-backed AI world, where founders typically dilute significantly at such scales. The fact that Liang was willing to commit personal capital to maintain control speaks to a deeply held vision for the company’s direction.

What the Money Buys

The $8 billion in new capital is expected to fund a massive expansion of DeepSeek’s compute infrastructure. Training next-generation foundation models requires extraordinary GPU clusters, and the company will need to secure access to tens of thousands of high-end accelerators—whether Nvidia H200s and B200s, or domestically produced alternatives like Huawei’s Ascend chips. Reports suggest DeepSeek is pursuing a dual-track hardware strategy, hedging against potential US export restrictions by building out both Nvidia-based and domestic accelerator clusters.

Beyond raw compute, the funding will support DeepSeek’s commitment to open-source AI development. Liang Wenfeng has repeatedly stated that open-source models are core to the company’s mission, a stance that has earned DeepSeek enormous goodwill within the global developer community even as it operates under the constraints of China’s regulatory environment. The V4-Flash release, with its MIT license and open weights, is a direct manifestation of that commitment.

The Bigger Picture: China’s AI Ascendancy in Usage Metrics

The OpenRouter data tells a story that extends well beyond DeepSeek. With Chinese models occupying nine of the top ten weekly slots, the platform’s usage data provides concrete evidence that Chinese AI labs are not merely catching up to their Western counterparts—they are increasingly dominating the global developer market in terms of actual deployment volume.

This shift is driven by the same formula DeepSeek has perfected: aggressive pricing, open-source releases, and architectures optimized for inference efficiency rather than raw training scale. While Western labs like OpenAI and Anthropic continue to compete on the absolute frontier of capability—and command premium pricing for it—the volume tier of the market is being claimed by Chinese models that offer 90% of the quality at 5% of the cost.

For enterprises and developers building production AI applications, the calculus is straightforward. When a model costs less than a tenth of a frontier alternative while delivering performance that is “good enough” for the vast majority of use cases, the economic pressure to switch becomes overwhelming. The 7.22 trillion tokens processed in a single week are the market voting with its wallet.

Looking Ahead

DeepSeek’s dual triumph—a globally dominant model and a massive funding round—positions the company as arguably the most consequential AI lab outside of the traditional Silicon Valley axis. The open question is whether the company can sustain its pace of innovation while scaling its infrastructure and navigating the complex geopolitical landscape that surrounds Chinese AI companies.

What is certain is that the narrative of AI supremacy being a two-horse race between OpenAI and Google is over. The center of gravity is shifting, and the OpenRouter rankings are the most honest reflection of where developers are actually placing their bets. DeepSeek V4-Flash’s 7.22 trillion tokens are not just a number—they are a signal that the global AI landscape has fundamentally changed.