Models · 2026-10-13
DeepSeek Model Lineup (October 2026): Every Model and Its Status
As of October 2026, DeepSeek's matrix spans language, reasoning, code, vision and math. Current lineup first, then legacy families.
Current API Models (October 2026)
| Model | Params | Highlights | Status |
|---|
| V4.1-Flash | 552B MoE | Native vision, off-peak halves, $0.003 cache | ✅ Newest (Sep 10; API name deepseek-flash) |
| V4 Pro (0813) | 1.6T / 49B active | Open-weight flagship, 1M context, MIT | ✅ Live (v4-pro routes to V4.1-Flash since Sep 14) |
| V4-Flash (0731) | 284B / 13B active | Cost tier, $0.14/$0.28 | ✅ Live |
V4 Family Timeline
- 2026-04-24: V4 family previewed (MIT license)
- 2026-07-31: V4-Flash-0731 official
- 2026-08-16: V4 Pro (0813) repriced
- 2026-09-10: V4.1-Flash launches, ~70% output cut
- 2026-09-14: v4-pro API calls route to V4.1-Flash
Legacy Models
| Model | Type | Notes |
|---|
| V3.2 / V3 | Language | Sparse attention; 671B MoE that reset open weights |
| R1 / R1-Distill | Reasoning | Reasoning series + six distilled models |
| Coder / Coder V2 | Code | Open-source MoE code models |
| VL / VL2 | Vision-language | Two generations of vision models |
| Janus | Multimodal | Unified understanding + generation |
| Math / Prover | Math | 7B math (GRPO); Lean 4 theorem proving |
| V2 / V2.5 | Language | First-generation MoE |
How to Choose
- General & multimodal:
deepseek-flash (V4.1-Flash) — newest, cheapest, sees images
- Heaviest reasoning: V4 Pro weights (self-host; API routing has switched)
- Code: Coder V2 or V4.1-Flash
- Self-hosting: V4 family is MIT; September's open-source Ascend stack runs it on domestic chips