Official announcements
Index of significant posts from the official DeepSeek site, newest first. Each links to the original.
We do not republish the posts.
Their copyright belongs to DeepSeek and they change. This is an index — title and date only.
Read the original.
| Date | Announcement | Link |
|---|---|---|
| 2026-09-10 | DeepSeek V4.1 Flash: stronger, faster, cheaper | 官方原文 |
| 2026-04-24 | DeepSeek-V4 Preview: the million-token context era | 官方原文 |
| 2025-12-01 | DeepSeek V3.2: stronger agents, reasoning integrated | 官方原文 |
| 2025-08-21 | DeepSeek-V3.1 released | 官方原文 |
| 2025-05-28 | DeepSeek-R1 update: deeper thinking, stronger reasoning | 官方原文 |
| 2024-12-26 | DeepSeek-V3 released | 官方原文 |
What they said (our read)
- V3 (2024-12) — training cost down to USD 5.57M, proving frontier models can be cheap.
- R1 (2025-01) — reasoning from pure RL, with six distills released at once. This started the ecosystem boom.
- V3.1 (2025-08) — thinking and non-thinking merged into one model.
- V3.2 (2026) — stronger agent capability, aimed at tool use.
- V4 (2026) — million-token context makes long documents routine.
- V4.1 Flash (2026-11) — faster and cheaper, for high-frequency calls.