Parameters
32B
Context Length
131K
Modality
Text
Architecture
Dense
KV Cache Rate
68 KB/tok
License
MIT License
Release Date
27 Dec 2024
Knowledge Cutoff
Jul 2024
API Pricing (per 1M)
Self-hosted only
Precision
Context
Workload Memory: 32.34 GB
INT4 · 8K context
Base Weights: 14.90 GB
KV Cache: 14.94 GB
Runtime & Graphs: 2.50 GB
1x NVIDIA A40 48GB
32.3 / 43.2 GB usable
2x RTX 4090 24GB
32.3 / 43.2 GB usable · TP 2
Apple M4 Max 128GB
32.3 / 96.0 GB usable
Determine requirements for your specific workload:
Open in VRAM Calculator
Checkpoint format
FP16
CUDA
>= 11.8
PyTorch
>= 2.1
Transformers
>= 4.43.1
| Architecture | Precision | Attention kernel |
|---|---|---|
Blackwell, Hopper | Native FP16 | FlashAttention-3 |
Ada Lovelace, Ampere | Native FP16 | FlashAttention-2 |
Rank
#173
| Benchmark | Score | Rank |
|---|---|---|
Graduate-Level QA | 0.621 | 95 |
Intelligence Index | 0.08 | 266 |
Overall Rank
#173
Coding Rank
-
DeepSeek-R1 32B is a high-capacity dense reasoning model distilled from DeepSeek-R1 onto the Qwen2.5-32B architecture. It offers frontier-level performance on complex coding, mathematical proofs, and analytical tasks with a 131K context window.
Attention
Attention Structure
Multi-Layer Attention
Attention Heads
96
Key-Value Heads
96
Position Embedding
ROPE
RoPE Theta
1,000,000
Sliding Window Attention
No
Sliding Window Size
131,072
Normalization
RMS Normalization
Activation Function
Swish
Dimensions
Base Precision
FP16
Hidden Dimension Size
8,192
Number of Layers
60
FFN Intermediate Size (Dense)
27,648
Tokenizer
Vocabulary Size
152,064
DeepSeek-R1 is a model family developed for logical reasoning tasks. It incorporates a Mixture-of-Experts architecture for computational efficiency and scalability. The family utilizes Multi-Head Latent Attention and employs reinforcement learning in its training, with some variants integrating cold-start data.