DeepSeek V3.2 / R2 (671B MoE)
Industry price-performance benchmark combining 671B total parameters with only 37B active parameters per token and ultra-compact Multi-Head Latent Attention (MLA).
Provider
DeepSeek (MIT Open-Weight)
Architecture
Fine-Grained MoE (671B Total / 37B Active) + Multi-Head Latent Attention
Parameters
671B Total / 37B Active
Context Window
256,000 tokens (Max Output: 65,536 tokens)
API Pricing
Input: $0.27 / 1M · Cached: $0.07 / 1M · Output: $1.10 / 1M
VRAM (16K Ctx)
Q4_K_M: 405.31 GB · Q8_0: 714.03 GB · FP16: 1345.27 GB
Hardware Target
8× H100/H200 (FP8) or Mac Studio M4 Ultra 192GB + Offload (IQ3_M)
Interactive Token Cost Estimator (DeepSeek V3.2 (671B))
GPU Hardware Compatibility Matrix (DeepSeek V3.2 (671B) at 16K Context)
| GPU / Hardware Profile | VRAM | Q4_K_M (4.8-bit) | Q8_0 (8-bit) | Est. Speed (Q4) |
|---|---|---|---|---|
| RTX 5060 Ti 16GB 448 GB/s | 16 GB | REQUIRES MULTI-GPU (405.31 GB) | REQUIRES MULTI-GPU (714.03 GB) | ~3 tok/s |
| RTX 4090 24GB 1008 GB/s | 24 GB | REQUIRES MULTI-GPU (405.31 GB) | REQUIRES MULTI-GPU (714.03 GB) | ~6 tok/s |
| RTX 5090 32GB 1792 GB/s | 32 GB | REQUIRES MULTI-GPU (405.31 GB) | REQUIRES MULTI-GPU (714.03 GB) | ~11 tok/s |
| M4 Max 64GB 546 GB/s | 52 GB | REQUIRES MULTI-GPU (405.31 GB) | REQUIRES MULTI-GPU (714.03 GB) | ~3 tok/s |
| 2× 24GB (48GB) 920 GB/s | 48 GB | REQUIRES MULTI-GPU (405.31 GB) | REQUIRES MULTI-GPU (714.03 GB) | ~6 tok/s |
| H100 80GB 3350 GB/s | 80 GB | REQUIRES MULTI-GPU (405.31 GB) | REQUIRES MULTI-GPU (714.03 GB) | ~21 tok/s |
| M4 Ultra 192GB 819 GB/s | 164 GB | REQUIRES MULTI-GPU (405.31 GB) | REQUIRES MULTI-GPU (714.03 GB) | ~5 tok/s |
| 8× H200 (1,128GB) 4800 GB/s | 1128 GB | FITS IN VRAM (405.31 GB) | FITS IN VRAM (714.03 GB) | ~142 tok/s |
Architectural Strengths & Target Workloads
- Unbeatable API economics ($0.07/1M cached input, $1.10/1M output) under permissive MIT license
- Only 37B active parameters yield high token/sec throughput on multi-GPU or hybrid CPU/GPU rigs
- Tiny MLA KV-cache (~70 MB per 1K tokens at FP16) allows high concurrency
Primary Engineering Workloads: High-Throughput Batch Code Analysis · Cost-Optimized CI/CD Agents · Self-Hosted Enterprise LLM.
Frequently Asked Questions (DeepSeek V3.2 (671B))
How much does DeepSeek V3.2 / R2 (671B MoE) cost per 1M tokens?
DeepSeek V3.2 / R2 (671B MoE) is priced at $0.27 per 1M uncached input tokens, $0.07 per 1M cached input tokens, and $1.10 per 1M output tokens (DeepSeek API (disk-cached prefix hits at $0.07 / 1M tokens) or Self-Hosted).
What is the context window of DeepSeek V3.2 / R2 (671B MoE)?
It supports up to 256,000 tokens (256,000 tokens) and a maximum generation output of 65,536 tokens.