Skip to content
NEWLive updates feed from Hugging Face, arXiv and GitHub  See what changed →

deepseek-ai · deepseek_v32 · based on deepseek-ai/DeepSeek-V3.2-Exp-Base

DeepSeek-V3.2

DeepSeek-V3.2 by deepseek-ai: a 685.4B-parameter mixture-of-experts open-weight model under the mit license.

mitMixture of experts164K contextAuto-imported · not yet reviewed

2,421,883 downloads · 1,490 likes · Model card · synced 2026-09-19

Specifications

Parameters
685.4B (21.4B active)
Architecture
moe · deepseek_v32
Layers
61
Hidden size
7,168
KV heads · head dim
128 · 56
Vocabulary
129,280
Native dtype
bfloat16
Context window
163,840 tokens
Released
2025-12-01
Last modified on Hub
2025-12-01

Features & licensing

License
mit
Commercial use
Yes
Modalities
text
Tool / function calling
No
Reasoning mode
No
Languages
EN
Quantised variants on Hub
AWQ, GGUF, bitsandbytes
Pipeline
text-generation
transformerssafetensorsdeepseek_v32text-generationeval-resultsendpoints_compatiblefp8

Task fit

Editorial scores (0–100) used by the recommendation engine.

  • Chat70
  • RAG / Q&A70
  • Code70
  • Summarisation70
  • Extraction70
  • Agents70

Hardware to run DeepSeek-V3.2

Weights need about 1508 GB at FP16, 792 GB at INT8 and 452 GB at INT4. The KV cache adds roughly 1749 MB per 1,000 tokens per request. For a reference workload of 5 requests per second with 1,500 input and 300 output tokens, total GPU memory is around 1253.1 GB.

GPU configurations at the reference workload
ConfigurationVRAMUtilisationEst. first tokenCloud / month

Estimates only. Run an assessment for your own traffic.

Auto-imported from Hugging Face. Architecture, license and feature data are synced from the model repository; the task-fit scores below are provisional defaults until a curator reviews them.