Latest Content
DeepSeek V3.2 Speciale icon

DeepSeek V3.2 Speciale

NVIDIA
DeepSeek-V3.2-Speciale is a reasoning-focused Mixture-of-Experts (MoE) large language model developed for advanced mathematical reasoning, coding, and complex analytical tasks. The model features a 685B parameter architecture with ~37B activated parameters, built on a 61-layer transformer with 128 attention heads and a 7,168 hidden size, utilizing 256 routed experts with 8 experts activated per token. It supports up to a ~160K token context window and integrates DeepSeek Sparse Attention (DSA) to reduce computational complexity while maintaining performance.
TypeMoE LLM
CapabilitiesText Generation, Instruction Following, Reasoning, Mathematical Reasoning+5 more
Links
Paper|Blog|HF Model Card
LicenseMIT

Inference Instructions

Deploy and run this model on NVIDIA B200 GPUs using the command below. Copy the command to get started with inference.

CONSOLE
docker run -it --rm 
 --runtime=nvidia 
 --gpus all 
 --ipc=host 
 --shm-size=128g 
 -p 8000:8000 
 -v ~/.cache/huggingface:/root/.cache/huggingface 
 -e HF_TOKEN='YOUR_HF_TOKEN' 
 -e LD_LIBRARY_PATH='/usr/local/nvidia/lib64:/usr/local/nvidia/lib:/usr/lib/x86_64-linux-gnu' 
 vllm/vllm-openai:v0.15.0-cu130 
 deepseek-ai/DeepSeek-V3.2-Speciale 
  --tensor-parallel-size 8 
 --max-model-len auto 
  --gpu-memory-utilization 0.90 
 --quantization fp8 
 --tokenizer-mode deepseek_v32 
 --reasoning-parser deepseek_v3 
 --max-num-seqs 1024 
 --disable-log-requests 
 --trust-remote-code

Model Benchmarks

Each model was tested with a fixed input size and total token volume while increasing concurrency to measure serving performance under load.

ITL vs Concurrency

Time to First Token

Throughput Scaling

Total Tokens/sec vs Avg TTFT

Vultr Cloud GPU

NVIDIA HGX B200

Deploy NVIDIA B200 on Vultr Cloud GPU

Deploy NVIDIA B200 instancesContact Sales