Can I Run DeepSeek Coder V2 236B MoE on NVIDIA DGX Spark (128GB Grace-Blackwell Unified)?
Full VRAM footprint analysis, tokens/sec generation benchmark, and copy-paste Ollama & vLLM runtime commands for running DeepSeek Coder V2 236B MoE (236B) on NVIDIA DGX Spark (128GB Grace-Blackwell Unified) (118 GB).
⚠️ Partial Offload — Slow CPU/RAM Spillover
OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 ollama run deepseek-coder-v2:236bLocal Hardware Configurations & Pricing
⚡ Partial Offload: Dedicated 100% GPU Options2x NVIDIA RTX 3090 Dual Rig (48GB Total VRAM)
Best price-to-VRAM homelab rig for 70B Q4_K_M & Qwen 3.8 27B Q8
Apple Mac Studio M4 Max (128GB Unified Memory)
Running Llama 4 Scout 109B & 70B Q8 locally with 128K+ context
NVIDIA DGX Spark (128GB Grace-Blackwell Unified)
128GB Unified Memory with native CUDA + Blackwell FP4 support
⚠️ Out of VRAM? Don't downgrade model quality.
Out of VRAMSpin up an on-demand cloud GPU instance in 60 seconds with pre-installed Ollama and vLLM. (DeepSeek Coder V2 236B MoE @ 151.34 GB VRAM)
Quantization Precision Matrix
32K ctx • KV: Q8_0| Quantization | Weights | KV Cache | Total VRAM | Est. Speed | Fit Status |
|---|---|---|---|---|---|
| FP1616 bpw | 472 GB | 4.03 GB | 493.3 GB (418%) | 0 t/s | OOM |
| Q8_08.5 bpw | 250.16 GB | 4.03 GB | 263.7 GB (223%) | 0 t/s | OOM |
| Q6_K6.56 bpw | 193.52 GB | 4.03 GB | 205.07 GB (174%) | 0 t/s | OOM |
| Q5_K_M5.68 bpw | 167.56 GB | 4.03 GB | 178.2 GB (151%) | 0 t/s | OOM |
| Q4_K_M4.8 bpw | 141.6 GB | 4.03 GB | 151.34 GB (128%) | ~4 t/s | PARTIAL |
| FP44.25 bpw | 125.08 GB | 4.03 GB | 134.24 GB (114%) | ~4 t/s | PARTIAL |
| IQ3_XS3.3 bpw | 96.76 GB | 4.03 GB | 104.93 GB (89%) | ~24 t/s | FITS |
Export Full Architecture & Infrastructure Spec
Get an executive-ready PDF report with exact memory math, hardware benchmarks, and cost comparisons.
Frequently Asked Questions: DeepSeek Coder V2 236B MoE on NVIDIA DGX Spark (128GB Grace-Blackwell Unified)
How much VRAM does DeepSeek Coder V2 236B MoE need on NVIDIA DGX Spark (128GB Grace-Blackwell Unified)?
At standard 4-bit quantization (Q4_K_M) with 32K context, DeepSeek Coder V2 236B MoE requires 151.34 GB of total VRAM (weights + KV cache + CUDA runtime overhead).
What inference speed will DeepSeek Coder V2 236B MoE achieve on NVIDIA DGX Spark (128GB Grace-Blackwell Unified)?
Based on NVIDIA DGX Spark (128GB Grace-Blackwell Unified)'s 273 GB/s memory bandwidth, expected decode speed is approximately 4 tokens per second.
Can I expand context beyond 32K tokens?
Each additional 32K tokens of context requires additional KV cache memory. For longer context windows, consider 8-bit or 4-bit KV cache quantization.