DeepSeek V4 Release Date: When It Launched and What Comes Next
DeepSeek V4 was officially released on April 24, 2026 — a date that sent ripples through the AI industry. The release came in the form of two preview models: DeepSeek-V4-Pro and DeepSeek-V4-Flash, both shipped simultaneously with open weights, API access, and a technical report.
This article covers the full DeepSeek V4 release timeline, what was announced, how the launch compared to previous DeepSeek milestones, and what you can expect from the road ahead.
The DeepSeek V4 Release Timeline
| Date | Event |
|---|---|
| March 9, 2026 | DeepSeek-V4-Lite (200B params) quietly released |
| April 24, 2026 | DeepSeek-V4-Pro & V4-Flash preview officially launched |
| April 24, 2026 | Open weights published on HuggingFace and ModelScope |
| April 24, 2026 | API updated (deepseek-v4-pro / deepseek-v4-flash) |
| July 24, 2026 | Legacy deepseek-chat & deepseek-reasoner retire |
The April 24 release was positioned as a preview, meaning DeepSeek may ship refined versions or expanded multimodal capabilities in future updates before the model reaches its final stable release.
What Dropped on April 24, 2026
The April 24 announcement was packed:
Two Model Tiers
- V4-Pro: 1.6T total / 49B active parameters — the flagship, targeting research, enterprise, and complex agentic tasks
- V4-Flash: 284B total / 13B active parameters — a leaner, faster, highly cost-effective option
Immediate API Availability
DeepSeek updated its API on launch day. Developers only needed to change the model field — no new SDKs, no integration work. The API maintains full compatibility with both OpenAI ChatCompletions and Anthropic API formats.
Open Weights on HuggingFace
Both models were published under the MIT License, enabling commercial use and local deployment. The Pro model requires approximately 865 GB storage; the Flash model requires ~160 GB.
Web Access via chat.deepseek.com
Instant Mode (powered by Flash) and Expert Mode (powered by Pro) went live the same day at chat.deepseek.com.
Why the Timing Matters
The April 24 launch came just weeks after OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 releases, making it a deliberate competitive response. DeepSeek V4 doesn't claim to beat these models on every benchmark — but it undercuts them dramatically on price while delivering near-frontier performance.
According to DeepSeek's own data, V4-Pro trails state-of-the-art closed models by roughly 3 to 6 months in developmental trajectory. Given how fast the AI landscape moves, that gap is remarkably narrow.
Before April 24: The DeepSeek V4 Lite Preview
Before the headline launch, DeepSeek quietly released DeepSeek-V4-Lite on March 9, 2026. This 200-billion-parameter model served as an early technical preview of V4's architecture innovations — particularly the Hybrid Attention mechanism (CSA + HCA) — allowing developers to begin testing ahead of the full release.
What's Next After the V4 Preview?
DeepSeek explicitly frames the April 24 release as a preview, which traditionally means:
- Stability improvements — resolving any inference quirks or edge cases found post-launch
- Expanded capabilities — potential multimodal additions (image, audio, video inputs)
- DeepSeek-V4.1 — following the V3 pattern, a V4.1 update is likely in the coming months
- Full stable release — the preview label may be dropped once evaluation is complete
AI platforms and developers watching this space — including tools like Framia.pro, which tracks and integrates frontier AI models for creative workflows — will be among the first to adopt new capabilities as they ship.
Legacy Model Retirement: Key Date
If you're currently using the old deepseek-chat or deepseek-reasoner API model names, take note: both will be permanently retired on July 24, 2026 (15:59 UTC). At launch, both legacy names were already routing to V4-Flash (non-thinking and thinking modes, respectively), so the transition is essentially seamless — but you should update your integrations before the deadline.
Summary
DeepSeek V4 launched on April 24, 2026, with two models (Pro and Flash), MIT-licensed open weights, immediate API availability, and a technical report detailing its Hybrid Attention Architecture. The release is currently in preview, with a stable version and potential expansions expected in the months ahead.
For anyone building with AI in 2026, the DeepSeek V4 launch is a key date to mark in history — it represents the moment that 1-million-token context became the default standard for an open-weight model.