DeepSeek 这家位于杭州的中国新创公司，发布了其新旗舰 AI 模型家族 V4 Flash 和 V4 Pro 的预览版，将其定位为迄今最强大的开源平台，并直接挑战 OpenAI、Anthropic、Google 和其他竞争对手。此次发布距离 DeepSeek 的 R1 模型震撼矽谷并因以低得多的成本展现前沿级表现而引发 1 兆美元抛售已超过 1 年。这次新版本强调架构与优化升级，包括一种旨在提升长上下文记忆的 Hybrid Attention Architecture，并将上下文窗口扩大到 1 million tokens，让整个程式码库或长文件可一次以单一提示送入。
DeepSeek 表示，V4 在程式码能力基准测试上达到顶尖水准，并在推理与 agentic 任务上有显著提升，同时仍专注于成本效率。该公司的 1 兆参数系统采用 Mixture-of-Experts，每个任务最多只启用 37 billion 参数，有助于维持较低的推理成本。对于 V4 Pro，输入 tokens 每 1 million tokens 收费 $1.74，输出 tokens 每 million 收费 $3.48；相比之下，Anthropic 的 Claude Sonnet 4 为每 million 输入 tokens $3、每 million 输出 tokens $15。DeepSeek 还表示，该模型在标准基准测试上优于 OpenAI 的 GPT-5.2，不过与最先进系统相比，仍落后约 3 到 6 个月。
此次部署受到算力紧缺的限制，DeepSeek 表示 V4 Pro 容量极为有限，但预计在今年下半年华为 Ascend 950 驱动的丛集上线后，价格将会下降。Bloomberg Intelligence 表示，该模型强化了中国以成本效率见长的 AI 声誉，但不太可能再引发一次撼动市场的「DeepSeek Moment」，并预期由于更容易取得先进 Nvidia 晶片，美国将保持约 6 个月的技术领先。这一消息带动 Semiconductor Manufacturing International Corp. 上涨 10%、Hua Hong Semiconductor 上涨 15%，而 Zhipu 下跌 9%；同时，这也发生在更广泛的 AI 支出之中，该支出预计将在 2026 年达到约 $650 billion，以及对于据称的蒸馏行为与中国可能使用被禁 Nvidia Blackwell 晶片的持续审查之下。
DeepSeek, a Hangzhou-based Chinese startup, unveiled preview versions of its new flagship AI model family, V4 Flash and V4 Pro, positioning it as the most powerful open-source platform yet and a direct challenge to OpenAI, Anthropic, Google, and other rivals. The launch comes more than 1 year after DeepSeek’s R1 model shook Silicon Valley and helped trigger a trillion-dollar selloff by showing frontier-level performance at much lower cost. The new release emphasizes architecture and optimization upgrades, including a Hybrid Attention Architecture designed to improve long-context memory, and expands the context window to 1 million tokens so entire codebases or long documents can be sent in a single prompt.
DeepSeek said V4 delivers top-tier coding benchmarks and major gains in reasoning and agentic tasks, while still focusing on cost efficiency. The company’s trillion-parameter system uses Mixture-of-Experts and activates only up to 37 billion parameters per task, helping keep inference costs low. For V4 Pro, input tokens cost $1.74 per 1 million tokens and output tokens cost $3.48 per million, versus Anthropic’s Claude Sonnet 4 at $3 per million input tokens and $15 per million output tokens. DeepSeek also said the model outperforms OpenAI’s GPT-5.2 on standard benchmarks, though it still trails state-of-the-art systems by about 3 to 6 months.
The rollout has been constrained by a computing crunch, with DeepSeek saying V4 Pro capacity is extremely limited, though it expects prices to fall after Huawei Ascend 950-powered clusters launch in the second half of this year. Bloomberg Intelligence said the model reinforces China’s cost-efficient AI reputation but is unlikely to spark another market-disrupting “DeepSeek Moment,” and expects the US to keep about a 6-month technical lead thanks to better access to advanced Nvidia chips. The news lifted Semiconductor Manufacturing International Corp. 10% and Hua Hong Semiconductor 15%, while Zhipu fell 9%; it also comes amid broader AI spending that is projected to reach around $650 billion in 2026, as well as ongoing scrutiny over alleged distillation and possible use of banned Nvidia Blackwell chips in China.