DeepSeek has released preview versions of its new DeepSeek V4 artificial intelligence models, putting the Chinese startup back at the center of the global AI race with a flagship system built around a 1 million-token context window and support for Huawei hardware. The launch includes DeepSeek V4-Pro, a 1.6 trillion-parameter mixture-of-experts model, and DeepSeek V4-Flash, a smaller 284 billion-parameter version designed to be faster and cheaper. The release also landed as tensions between Washington and Beijing sharpened, with the White House accusing Chinese entities of trying to steal American AI technology and China rejecting those claims.
DeepSeek is presenting V4 as a lower-cost, open-source model family aimed at advanced reasoning, coding, and agent-style tasks. The company said V4 supports an ultra-long context length of one million tokens, which determines how much text a model can absorb while completing a task, and it described V4-Flash as the more economical option between the two releases. OfficeChai reported that both models are available through API access and through DeepSeek’s web and app products, with V4-Pro offered in Expert Mode and V4-Flash in Instant Mode.
Pricing and rollout
Pricing is a major part of DeepSeek’s pitch for V4. One source reported that V4-Pro costs $3.48 per million output tokens, while V4-Flash costs $0.28 per million output tokens, placing both well below the output pricing cited for GPT-5.4 and Claude Opus 4.6. DeepSeek also described the new model as having drastically reduced compute and memory costs, reinforcing the company’s long-running strategy of competing on price as well as performance.
Hardware is another reason the launch is drawing attention. Tom’s Hardware described V4 as the first major frontier AI release optimized for Huawei’s Ascend processors rather than Nvidia hardware, and DeepSeek said the model can run on chips made by Huawei. Huawei said its full range of Ascend SuperPoD products is supporting the DeepSeek V4 series, a sign that Chinese AI developers are moving to build stronger domestic alternatives in advanced computing infrastructure.
Benchmarks and capabilities
DeepSeek says V4-Pro is strongest in agent capabilities, world knowledge, and reasoning performance inside China, while also claiming that the model trails only the latest Gemini system on one world knowledge benchmark. OfficeChai’s review of the released benchmarks said V4-Pro matched GPT-5.4 on MMLU-Pro at 87.5, led on LiveCodeBench with a score of 93.5, and posted a Codeforces rating of 3206, ahead of GPT-5.4 and Gemini-3.1-Pro. The same report said V4-Pro remained behind top closed-source rivals on some tests, including Humanity’s Last Exam and SimpleQA-Verified, showing that DeepSeek’s latest model is highly competitive but not clearly dominant across every category.
V4-Flash appears to be the budget model without being a stripped-down afterthought. OfficeChai reported that the gap between Flash and Pro is usually around one to three points on many benchmarks, although the lighter model drops more sharply on tests tied to complex tool use and detailed factual recall. Both models share a 1 million-token context window, support thinking and non-thinking modes, and offer features such as JSON output and tool calls, while V4-Pro is also positioned for use with coding and agent products, including Claude Code, OpenClaw, OpenCode, and CodeBuddy.
China AI rivalry
The V4 release arrives more than a year after DeepSeek shook the AI industry with a lower-cost reasoning model that challenged assumptions about U.S. leadership in the field. NDTV reported that DeepSeek’s earlier rise helped trigger a sell-off in AI-related shares and fueled fresh debate over how much computing power is really needed to build advanced models. OfficeChai also said DeepSeek’s R1 model once drew enough attention to briefly top the U.S. App Store ahead of ChatGPT, though the company later faced stronger competition from other Chinese labs, including Alibaba’s Qwen.
At the same time, politics now sit much closer to the product story. Tom’s Hardware reported that the U.S. State Department sent a diplomatic cable warning foreign governments about alleged intellectual property theft by Chinese AI firms, while Michael Kratsios, President Donald Trump’s science and technology adviser, said the United States had evidence of industrial-scale distillation campaigns based mainly in China. China’s foreign ministry called those accusations groundless, leaving DeepSeek’s V4 debut as both a technology launch and a fresh flashpoint in the broader U.S.-China fight over AI leadership, chip access, and model development.
