On April 24, 2026, DeepSeek V4 has officially surfaced. Unlike a few hours ago, when it could only be judged based on media rumors, now DeepSeek's official Hugging Face organization has launched DeepSeek-V4-Pro, DeepSeek-V4-Flash, and the corresponding Base model, and the official API price page also appears at the same time. This means that DeepSeek V4 is no longer just a "rumor to be released", but has entered the public verifiability stage.
The focus of this release is clear: DeepSeek V4 focuses on million-token long contexts, open-source weighting, MoE architecture, and lower long-context inference costs. What it means for developers is not just "another strong model", but also brings new options for long documentation, codebases, agent workflows, and on-premises/privatized deployments.
Two product lines: Pro and Flash
According to the official model card, the DeepSeek V4 series contains two main versions: DeepSeek-V4-Pro with a total parameter of 1.6T and an activation parameter of 49B; DeepSeek-V4-Flash with a total parameter of 284B and an activation parameter of 13B. Both support 1M context length and provide Base and instruction model weights.
From the perspective of positioning, Pro is the flagship version, aimed at complex inference, code, agents, and long context tasks; Flash is more speed-oriented and cost-oriented, suitable for regular Q&A, batch processing, and latency-sensitive applications. Officials also mentioned that V4-Pro-Max is Pro's highest inference effort mode, with the goal of raising the upper limit of knowledge, code, and agent tasks in open source models.
API prices and interfaces are also synchronized
DeepSeek's official API pricing page already lists the V4 model: deepseek-v4-flash and deepseek-v4-pro both use https://api.deepseek.com as the OpenAI format base URL, and also provide an Anthropic format base URL. Both support thinking and non-thinking modes, JSON output, tool calls, Chat Prefix Completion, and FIM Completion is only available in non-thinking mode.
In terms of price, deepseek-v4-flash hits $0.028 per million input token cache hits, $0.14 cache misses, and $0.28 outputs; deepseek-v4-pro corresponds to $0.145, $1.74, and $3.48. The old names deepseek-chat and deepseek-reasoner will be deprecated in the future, but for compatibility, they correspond to deepseek-v4-flash's non-thinking mode and thinking mode respectively.
Technical highlights: Long-context costs are declining
The DeepSeek V4 model card focuses on long-context efficiency with core upgrades. Officially, V4 uses a mixed attention architecture, combining Compressed Sparse Attention and Heavily Compressed Attention. In the context of 1M tokens, DeepSeek-V4-Pro only requires 27% single-token inference FLOPs and 10% KV cache compared to DeepSeek-V3.2.
This is critical for real-world applications. In the past, millions of contexts often sounded strong, but cost, memory, and latency make it difficult to use normally. If V4's long-context efficiency is consistent in community deployments and cloud APIs, it will directly impact code repository understanding, legal contract review, scientific research collation, enterprise knowledge base, and multi-round agent tasks.
What to look at now
The most noteworthy thing about the release of DeepSeek V4 this time is not the single list number, but the combination of "open weight + API availability + 1M context + low-price cache". For domestic developers, it will continue to drive the adaptation of domestic AI chips and inference frameworks; for the global open source community, V4-Pro and V4-Flash will become new benchmarks for evaluation, quantification, deployment, and agent framework access.
However, it is still necessary to maintain a little caution when landing: when the large model is first released, third-party running scores, actual throughput, tool call stability, and adaptation of different inference frameworks still need time to verify. In the production environment, you can do A/B testing first, and do not directly replace existing models based on parameter scale alone.
Information source: DeepSeek-V4-Pro official model card, DeepSeek official Hugging Face organization page, DeepSeek API model and price page, DeepSeek official public account release link.
