DeepSeek quietly pushed something new, and I almost missed it

Hello, my friends. Anna here. I opened the app this morning expecting the same 128K context window I'd been working with all week. Instead, I got a notification to update to version 1.7.4.

DeepSeek started a gray-scale rollout of what might be their V4 model, or at least V3's final form before V4 drops. And unlike most AI updates that happen quietly in the background, this one you can force if you know the trick.

The manual upgrade workaround

Someone on social media figured out the pattern. On the web version, if you ask "what's your context length," DeepSeek tells you 128K tokens with a knowledge cutoff of July 2024. That's the old version.

But if you update the mobile app to 1.7.4(1) and ask the same question, suddenly you're talking to a model with 1 million tokens of context and knowledge updated through May 2025. Then—and this is the part that caught my attention—when you go back to the web version and ask again, it's also upgraded. The gray-scale test apparently syncs across your account once triggered through the app.

I tested this. It works. The web version now shows the same expanded capabilities without any additional steps.

What actually changed

The jump from 128K to 1M tokens isn't just a bigger number. According to reports on DeepSeek's gray-scale testing, this represents nearly a tenfold increase in context capacity. For long documents, multi-day project planning, or complex ongoing work, this matters.

The knowledge cutoff moved from July 2024 to May 2025—ten months of additional training data. When I asked about late 2024 and early 2025 events, the updated model knew about them. The old version hallucinated or admitted ignorance.

Core capabilities improved noticeably, though DeepSeek hasn't published official benchmarks. Responses felt more coherent across longer conversations.

The V4 question nobody's answering clearly

Is this actually V4? DeepSeek's own responses suggest it's not—at least not officially. When users asked the updated model directly, it indicated this might be "the final evolutionary form of the V3 series" or "the ultimate gray-scale version before V4's official release," as reported by Futunn's analysis.

That phrasing matters. If this is V3's final form, it means V4 is still coming, possibly mid-February 2026 according to industry reports. If it's a gray-scale test of V4, then this update is a preview of what's next.

Either way, the practical result is the same: the model available right now is substantially better than what was available last week.

Why this rollout method is unusual

Most AI companies push updates automatically. DeepSeek's approach—requiring a mobile app update to trigger the gray-scale test—feels oddly deliberate. It means adoption will be uneven. Users who never open the mobile app won't trigger the upgrade.

I'm not sure if this is intentional strategy or just how gray-scale testing works at DeepSeek.

What I'm watching for

The knowledge cutoff update to May 2025 positions DeepSeek between V3's January 2025 cutoff and current frontier models—a tighter lag than most open-weight models maintain.

The 1M context window puts it in the same category as Claude's extended context, though implementation details matter more than raw numbers. How well does it maintain coherence across that full context? I'll need more time with complex tasks to know.

For now, the updated version handles long project planning sessions without losing track of earlier constraints. The knowledge refresh means I'm not constantly correcting outdated information. And forcing the upgrade myself, rather than waiting for DeepSeek to push it, felt unexpectedly satisfying.