DeepSeek, the Chinese AI startup that shocked global markets last year, just released preview versions of its V4 model series with a breakthrough capability: the ability to process 1 million tokens (roughly 750,000 words) in a single pass, while dramatically cutting the computing power and memory required to do so. The new models come in two versions, V4-Pro for high-performance tasks and V4-Flash for cost-effective deployment, and represent what industry experts are calling an "inflection point" for the entire AI industry.

To put this in perspective, 1 million tokens means DeepSeek's model could process all three books of the "Three-Body Problem" trilogy simultaneously without breaking a sweat. This is a massive jump from V3's previous 128,000 token limit. The company achieved this through a novel token-dimension compression mechanism combined with DSA sparse attention technology, which dramatically reduces computational and memory requirements compared to traditional methods.

How Does DeepSeek V4 Compare to Other Leading AI Models?

DeepSeek's new models are competitive with the world's most advanced AI systems. On knowledge benchmarks, V4-Pro leads all open-source models and trails only Google's proprietary Gemini-Pro-3.1. In mathematics, science, and competitive coding tests, it surpasses every open-source model with published results. The model was trained on an unprecedented 14.8 trillion tokens, giving it broad knowledge across many domains.

The agentic coding capabilities represent the most significant leap from previous versions. V4-Pro has become the default coding model for DeepSeek's internal engineering team, and internal evaluations show it outperforms Claude Sonnet 4.5 and delivers quality approaching Claude Opus 4.6's non-reasoning mode, though it still lags on complex reasoning tasks. V4-Flash matches the Pro version on simpler agent tasks and offers comparable reasoning for everyday queries.

What Makes the Cost Reduction So Significant?

One of the most striking aspects of V4's release is the pricing structure, which remains unchanged from previous versions despite the dramatic capability improvements. V4-Pro costs $2.19 per million input tokens and $7.19 per million output tokens, while V4-Flash is significantly cheaper at $0.27 per million input tokens and $1.07 per million output tokens. For context, this means processing a 100,000-word document would cost roughly $0.22 with Flash or $2.19 with Pro.

"This addresses the long-standing issues of slower performance and higher costs associated with long context lengths, marking a genuine inflection point for the industry," said Zhang Yi, founder of tech research firm iiMedia.

Zhang Yi, Founder, iiMedia

The cost reduction is particularly important for practical applications. Long-text processing, which previously required expensive high-end research labs, is now expected to move into mainstream commercial applications. This could transform how businesses handle document analysis, legal review, medical records processing, and other text-heavy workflows.

How to Understand DeepSeek V4's Technical Specifications

• Model Parameters: V4-Pro has 1.6 trillion parameters (the adjustable weights that refine decision-making), while V4-Flash has 284 billion parameters, making it more efficient for cost-conscious users
• Context Window: Both versions support 1 million tokens, matching Google's Gemini and putting them on par with the longest-context models available
• Hardware Compatibility: The models have been optimized to run on Huawei's Ascend chips, representing a strategic move toward technological independence from US-based semiconductor suppliers
• Agentic Optimization: V4 has been specifically optimized for AI Agent products including Claude Code, OpenClaw, OpenCode, and CodeBuddy

The hardware compatibility with Huawei's Ascend SuperPoD products is particularly noteworthy. DeepSeek had to completely rewrite core code to migrate from Nvidia's CUDA ecosystem to Huawei's CANN architecture, a technically challenging process that signals deepening technological decoupling between China and the United States. This move allows DeepSeek to operate independently of US export controls on advanced semiconductors.

What Does This Mean for the Global AI Competition?

DeepSeek's V4 release comes amid intensifying tensions between China and the United States over AI technology. The White House accused Chinese entities of running "industrial-scale distillation campaigns" to steal American AI technology, a claim Beijing rejected as "baseless". Distillation is a common practice in AI development where companies create cheaper, smaller versions of their own models.

"This is no less shocking than when DeepSeek first came out if its new model indeed matches the performance of leading models from Western labs," noted Max Liu, a veteran AI industry analyst.

Max Liu, AI Industry Analyst

DeepSeek's strategy of releasing open-source models, where the inner workings are publicly available, contrasts sharply with the proprietary approach taken by OpenAI and other Western rivals. This open-source strategy has driven widespread adoption among Chinese municipalities, healthcare institutions, and financial sector businesses. The combination of competitive performance, lower costs, and open-source availability positions DeepSeek as a serious challenger to Western AI dominance.

The V4 release also marks a milestone for China's domestic AI industry. By demonstrating that Chinese companies can match or exceed Western capabilities while operating under US export restrictions, DeepSeek is proving that technological self-reliance is achievable. This success is expected to accelerate competition in the domestic AI market and encourage more Chinese firms to develop advanced models.

DeepSeek announced that 1 million token context windows will become standard for all its official services going forward, signaling that this capability is no longer a premium feature but an industry baseline. A preview version of the open-source model is now available, though the company has not yet indicated when a final version will be released.