Parameters
14B
Context Length
131K
Modality
Text
Architecture
Dense
KV Cache Rate
45 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: 15.27 GB
INT4 · 8K context
Base Weights: 6.52 GB
KV Cache: 6.25 GB
Runtime & Graphs: 2.50 GB
1x NVIDIA L4 24GB
15.3 / 21.6 GB usable
2x RTX 4090 24GB
15.3 / 43.2 GB usable · TP 2
Apple M4 24GB
15.3 / 18.0 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
#180
| Benchmark | Score | Rank |
|---|---|---|
Graduate-Level QA | 0.591 | 96 |
Intelligence Index | 0.08 | 261 |
Overall Rank
#180
Coding Rank
-
DeepSeek-R1 14B is a dense reasoning model distilled from DeepSeek-R1 onto the Qwen2.5-14B backbone. It delivers competitive multi-step mathematical reasoning and programming synthesis across a 131K token context window.
Attention
Attention Structure
Multi-Layer Attention
Attention Heads
80
Key-Value Heads
80
Position Embedding
ROPE
RoPE Theta
1,000,000
Sliding Window Attention
No
Sliding Window Size
131,072
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Base Precision
FP16
Hidden Dimension Size
5,120
Number of Layers
40
FFN Intermediate Size (Dense)
13,824
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.