DeepSeek Silently Updates: 1M Token Context Unlocked, A Prelude to V4?

[Core Summary] DeepSeek has completed a major update on its web and app platforms, officially supporting 1M (1 million) Token ultra-long context. This move is not only a technical preview for the upcoming DeepSeek V4 in mid-March but also marks the entry of open-source/high-cost-performance models into the million-level context era, competing directly with Google Gemini 3 Pro (2M) and Anthropic Claude 4.5 Opus (1M).

---

[In-Depth Interpretation]

1. Closing the Gap, Chasing the First Tier

Previously, DeepSeek V3 supported a maximum of 128k context. This update to 1M means it has finally caught up with the specifications of Claude 4.5 Opus (1M) and Gemini 3 Pro (2M) in terms of long-text processing capabilities.

• Use Cases: Analyze entire novels, refactor large codebases, audit cross-year financial reports.

2. Core Highlight: Cost-Performance & Inference Speed

While Claude 4.5 and Gemini 3 Pro still lead in context window size and reasoning capabilities (especially in multimodal and complex logic), DeepSeek's core killer feature remains "Extreme Cost-Performance". If DeepSeek V4 can maintain V3's low-cost advantage (Input: $0.14/1M vs Claude 4.5's $15+/1M) under 1M context, it will create a huge commercial impact on closed-source models, especially in Long-Context RAG and Code Completion scenarios.

3. V4 is Poised to Launch

The industry widely predicts that DeepSeek V4 will be officially released in mid-March 2026 (around the Spring Festival). Today's 1M update is highly likely a grey-scale implementation of one of its core features (DeepSeek Sparse Attention, DSA). Currently, the API side has not fully opened the 1M limit, so developers are advised to pay close attention.

[Experience Entry]

• Web Experience: chat.deepseek.com
• App Download: Search "DeepSeek" in major app stores

[More Information]

For the DeepSeek V4 release countdown, latest benchmarks, and API access guides, please visit: 👉 deepseekv4.app