DeepSeek built the most permissive frontier model of its generation, then quietly locked away the one feature users were waiting for. That is the defining paradox of DeepSeek V4—and it is already shaping how developers talk about the model.
Engram, the long-teased memory and context-extension system that the community expected to arrive alongside V4, is nowhere in the base release. Instead, DeepSeek has bundled it into a premium tier, accessible only through the company's API at higher price points than the open weights download. For developers who celebrated V4's fully open architecture, this feels like a bait-and-switch wrapped in open-source branding.
What makes this especially frustrating is that Engram was not a rumor. DeepSeek researchers had discussed the capability in public interviews and conference talks throughout 2025, positioning it as the feature that would make V4 genuinely useful for long-horizon tasks—legal document analysis, full-codebase reasoning, extended research sessions. The community built expectations around it. Tutorials appeared before the model did.
Now those same developers are left with a powerful model that caps out around 128K context in practice, despite benchmark scores that suggest it should handle more. Some are turning to third-party context-window extensions, which work but introduce latency and reliability issues that defeat the purpose of using a high-performance model in the first place.
The competitive picture makes this stranger still. Anthropic released Claude 3.7 with native extended thinking and a context window that genuinely performs at 200K. OpenAI has steadily expanded what o3 can hold in active memory. Both companies charge for their premium tiers, but they do not hide core capability behind those tiers—users know what they are buying. DeepSeek V4's open weights are impressive on paper, but the moment a developer needs Engram for production use cases, they are paying DeepSeek API rates anyway, without the transparency of knowing exactly what they are purchasing.
For the open-source community, this raises a governance question that goes beyond one product cycle. DeepSeek benefits enormously from the goodwill generated by releasing powerful base models under permissive licenses. That goodwill is a form of currency. Locking core features behind proprietary tiers—features the community was led to expect as part of the V4 generation—depletes that currency faster than the benchmark scores suggest.
The pricing itself is not the scandal. Model providers need sustainable business models. The scandal is the silence. DeepSeek has not published documentation explaining what Engram does, how it works, or why it landed in a premium tier rather than the open release. That vacuum has filled with speculation: some developers assume it is technically unstable; others assume it is simply a monetization play; a few argue the feature never existed in the form users imagined.
What is certain is that DeepSeek V4 ships incomplete for anyone building serious production applications. The open weights remain genuinely useful for research and experimentation. But for the developers who planned their 2026 infrastructure around Engram, the release is a quiet disappointment—not because the model is bad, but because it is less than it promised to be.