DeepSeek Silently Expands Context Window 10x to One Million Tokens

DeepSeek quietly updated its web application on February 11, 2026, expanding the context window from 128,000 tokens to one million tokens — a nearly 10x increase that enables processing of entire codebases, legal document sets, or research corpora in a single session.

What Changed

The expansion applies to DeepSeek's web application and interface only; the API service remains on the V3.2 model with the previous context limit. DeepSeek also updated its knowledge cutoff to May 2025, bringing the model's training data closer to the present. The update was not accompanied by a formal announcement or blog post — users discovered the expanded context window through the interface itself, a pattern consistent with DeepSeek's historically understated approach to product updates.

What a Million Tokens Enables

A one-million-token context window can hold approximately 750,000 words of text — equivalent to roughly 10 average-length novels, several hundred pages of legal contracts, or the complete source code of a medium-to-large software project. For developers, this means that an entire codebase can be loaded into a single conversation, allowing the model to understand cross-file dependencies, architectural patterns, and system-level interactions that are invisible when working with individual files or small code snippets.

For researchers and analysts, the extended context enables processing of complete datasets, multi-paper literature reviews, and full regulatory documents without the summarization or chunking strategies that shorter context windows require. These workarounds — splitting documents into smaller pieces and processing them sequentially — introduce information loss and are a significant source of errors in current AI-assisted analysis workflows.

Competitive Context

Google's Gemini models introduced million-token context windows in 2024, and Anthropic's Claude models support 200K tokens with extended context available in enterprise plans. DeepSeek's expansion narrows the context window gap between Chinese and Western models, and does so in a consumer-facing product rather than an enterprise-only API — making long-context capabilities available to individual users without enterprise pricing.

The timing suggests preparation for DeepSeek V4, which is expected to feature a trillion-parameter architecture with particular strength in coding tasks. The context window expansion may be an incremental deployment of infrastructure capabilities that will ship as part of the V4 release, with the web application serving as a testing ground for the expanded context processing pipeline.