DeepSeek-V3.2
Canonical page on the main site: sinoaihub.com/models/deepseek-v3-2
Provider
Model Family
DeepSeek-V3
Release Date
2025-12-01
Status
discontinued
Open Weight
Yes
License
MIT
Self Hosting
Yes
Api Available
No
Official Api
No
Related Technologies
Architecture
DeepSeek Sparse Attention (DSA) MoE — 61 layers, hidden size 7168, MoE intermediate 2048, 256 routed experts (8 selected + 1 shared expert), vocab 129,280 (from config.json)
Known Limitations
- Replaced on the DeepSeek API by the V4 family
- Context window not stated on the official release pages fetched
- Vendor performance claims are qualitative ('GPT-5 level performance'); no numeric scores on the release page
Last Verified
2026-09-20
Sources
| source_name | source_url | source_type | published_date | last_verified | confidence |
|---|---|---|---|---|---|
| DeepSeek-V3.2 release | https://www.deepseek.com/en/news/deepseek-v3-2/ | official | 2025-12-01 | 2026-09-20 | high |
| DeepSeek Transparency Center | https://www.deepseek.com/en/transparency/ | official | — | 2026-09-20 | high |
| Hugging Face model card — DeepSeek-V3.2 | https://huggingface.co/deepseek-ai/DeepSeek-V3.2 | official | — | 2026-09-20 | high |