DeepSeek timeline
Every significant release from founding to now. Dates follow official announcements; where only the year is confirmed we say so.
| Date | Event | What changed | Source |
|---|---|---|---|
| 2023-07 | Company founded | Incubated by High-Flyer, based in Hangzhou. | — |
| 2023-11 | DeepSeek Coder / DeepSeek LLM 67B | First open-weight releases. | 来源 |
| 2024-02 | DeepSeekMath | Math reasoning model, introduced GRPO. | 来源 |
| 2024-05 | DeepSeek-V2 | Introduced MLA plus MoE, cutting inference cost. | 来源 |
| 2024-06 | DeepSeek-Coder-V2 | Code model covering many languages. | 来源 |
| 2024-12-26 | DeepSeek-V3 | 671B MoE, 37B active, trained for about USD 5.57M. | 来源 |
| 2025-01 | DeepSeek-R1 | Reasoning model with long chain-of-thought from RL; six distills released. | 来源 |
| 2025-05 | DeepSeek-R1-0528 | R1 update, deeper reasoning. | 来源 |
| 2025-08-21 | DeepSeek-V3.1 | Hybrid thinking and non-thinking modes. | 来源 |
| 2026 | DeepSeek-V3.2 | Stronger agent capability with reasoning integrated. | 来源 |
| 2026 | DeepSeek-V4 Preview | One million token context. | 来源 |
| 2026-11 | DeepSeek-V4.1 Flash | Faster, cheaper Flash line update. | 来源 |
How to read it
Three phases:
- 2023 — foundations. Coder and LLM first, proving it could train.
- 2024 — cost. MLA plus MoE in V2 cut inference cost; V3 cut training cost.
- 2025–2026 — capability. R1 reasoning, V3.1 hybrid modes, V3.2 agents, V4's million-token context.
One through-line: each generation turns the previous generation's capability into a baseline, and drops the price.