DeepSeek V3.2
#30 en Modèles de langage open sourceDeepSeek · v3.2 · depuis 2025-12-01 · 37× · vu le 14 août 2026
37
Momentum
DeepSeek V3.2 is an open-source language model (MoE architecture) by DeepSeek with 685 billion total parameters (671B main model + 14B MTP module), activating only ~37B parameters per token. Released on December 1, 2025, it introduces DeepSeek Sparse Attention (DSA), a novel fine-grained sparse attention mechanism that substantially reduces computational complexity and KV cache usage in long-context scenarios. Designed as an "agent-first" model, it is the first in the series to integrate thinking (chain-of-thought reasoning) directly into tool-use, supporting both thinking and non-thinking modes. Model weights and code are available under the MIT License.
Historique du momentum
19.05.17.08.
Fonctionnalités
| Context Window (Tokens) | 163,840 tokens (max context length; per official tech report: 128K native, extended to ~163K) |
| License | MIT License (model weights and code) |
| Multimodality | Text only (no native image/audio/video input); dedicated multimodal model is DeepSeek-VL2 (separate product) |
| Platform | DeepSeek Web, App, API (deepseek-chat); model weights on Hugging Face & GitHub; supported inference frameworks: SGLang, vLLM (Day-0 support); also available on Microsoft Azure Foundry |
| Price | Free to use via DeepSeek Web & App; API: from $0.028/1M input tokens (cache hit) to $0.28/1M input tokens (cache miss), $0.42/1M output tokens |
| Price per 1M Tokens | $0.28 / 1M input tokens (cache miss); $0.028 / 1M input tokens (cache hit, 90% discount); $0.42 / 1M output tokens – via DeepSeek Official API |
| Release Date | December 1, 2025 |