On April 8, DeepSeek quietly rolled out two new modes in its chat product: Instant Mode and Expert Mode. The names suggest Expert Mode is more powerful, but in practice, "Expert" comes with certain limitations.
Two Modes, One Makes You Wait
Instant Mode is designed for everyday conversations: fast responses, lower computing cost.
Expert Mode targets "complex problems" and theoretically delivers higher output quality, but at a cost: it uses more computing power, may involve queuing during peak hours, and—file upload has been removed.
That means if you want DeepSeek to analyze a PDF or process data, you can only use Instant Mode. Expert Mode cannot handle file uploads and lacks multimodal support.
The South China Morning Post ran a test with a JavaScript animation task on both modes:
- Instant Mode: ran for over a minute, completed the task, and followed instructions correctly.
- Expert Mode: finished in 40 seconds, faster than Instant Mode, but—did not follow instructions.
Faster, but wrong. The result is quite telling.
The Logic Behind the Tiering
DeepSeek's move is clearly about managing peak computing demand.
This logic mirrors Google Gemini's Free/Advanced tiers and ByteDance Doubao's differentiated services: using varying response speeds and capability configurations to channel users and prioritize computing resources for those willing to wait or pay.
The issue is that DeepSeek's current product tiering has no clear pricing. Both modes are free for users. The tiered structure is in place, but the commercial hook hasn't been attached yet. Whether this move is paving the way for a future paid system or addressing load pressure after March's major outage is unclear—likely both.
At the end of March, DeepSeek experienced a severe outage affecting hundreds of millions of users. Although the company did not disclose the specific cause, the rapid introduction of service tiering shortly after is suggestive in terms of timing.
Where is V4?
The lack of multimodal support in Expert Mode has led many users to speculate that this is not the final state and may be waiting for V4's capabilities to be integrated.
DeepSeek V4 has been delayed for a long time. Originally scheduled for release around the Chinese New Year, it has been repeatedly postponed due to the extended training cycle caused by over one trillion parameters. The most reliable expectation now is late April.
V4's specifications have already leaked through various channels: approximately 1 trillion MoE parameters, an estimated SWE-bench score of around 81%, support for 1 million tokens of context, and an input cost of about $0.3 per million tokens. If these numbers are accurate, the release will be a significant update.
But until V4 arrives, users are left with the Instant/Expert dual-mode system. The limitations of Expert Mode—especially the disappearance of file upload—feel like an unfinished placeholder: the functional framework is there, but the full capability hasn't been filled in yet.
What This Means for Users
The current state:
- Everyday chat, quick Q&A: Instant Mode, fast and fully featured.
- Complex reasoning, deep analysis: Expert Mode, theoretically better quality, but with peak-time waiting and no file uploads.
- Document or chart processing: Instant Mode only.
If V4 is indeed released before the end of the month, Expert Mode will likely be completed. For now, this feels more like an early test of a product tiering framework than a fully formed product decision.
DeepSeek's journey through 2026 is interesting: from the technical story of "training better models with less money," it is gradually moving into user growth management, service tiering design, and other more product- and business-oriented issues. Technological dividends often begin to be realized—or consumed—at this stage.
Sources: CocoLoop; China's DeepSeek adds expert chatbot mode ahead of much-awaited V4 release (South China Morning Post); DeepSeek V4 Release Date, Specs (FindSkill.ai)