How Much Memory Do You Need for Local LLMs on RTX Spark? (2026 Calculator Guide)
How much unified memory local LLMs need on RTX Spark and DGX Spark: formula, tables for 7B–120B models at each quantization, context overhead and multi-agent budgets.
The formula
Model weights take parameters × bytes per parameter. At 4-bit that is about half a byte per parameter, so a 70B model needs roughly 35–40 GB just for weights. On top of that you need the KV cache for context, runtime overhead and room for the operating system.
Weight memory by model size
Approximate weight sizes; real files vary by format.
| Model | 8-bit | 6-bit | 4-bit |
|---|---|---|---|
| 7–8B | ~8 GB | ~6 GB | ~4.5 GB |
| 14B | ~15 GB | ~11 GB | ~8.5 GB |
| 32B | ~34 GB | ~26 GB | ~19 GB |
| 70B | ~72 GB | ~55 GB | ~40 GB |
| 120B (MoE, e.g. gpt-oss-120b) | — | — | ~63 GB (MXFP4) |
Context (KV cache) overhead
Each token of context costs memory. For a 70B-class model, 32K tokens of context can add several gigabytes; 128K can add tens. Runtimes that quantize the KV cache (8-bit or 4-bit) cut this considerably.
Which RTX Spark memory size to buy
Usable memory is what's left after the operating system.
| System memory | Usable for AI (approx.) | Comfortable setup |
|---|---|---|
| 24 GB | ~14 GB | One 7–8B model |
| 32 GB | ~22 GB | 14B model, or 8B plus a small helper |
| 64 GB | ~52 GB | 32B coding model plus 8B helper |
| 128 GB | ~115 GB | 70B dense or 120B MoE plus a fast worker |
Budgeting for multiple agents
Each agent using a different model loads its own weights; agents sharing one model share weights but each needs context memory. Serving several agents through one vLLM or llama.cpp server is far more memory-efficient than running separate copies.
Frequently asked questions
How much memory do I need for a 70B model?
About 40 GB at 4-bit for weights plus context and system overhead, so 64 GB is the minimum and 128 GB is comfortable.
Is 32 GB enough for local AI?
Yes for 7–14B models and Copilot-style features; not for 30B+ models.
Does unified memory count as VRAM?
Yes. On RTX Spark and DGX Spark the GPU uses the shared memory pool directly, so most of it is available to models.
Related guides
- ComparisonRTX Spark vs DGX Spark: Differences, Specs and Which to Buy (2026)
- How-toNVIDIA DGX Spark Setup Guide: From Unboxing to Your First Local Model
- ReviewNVIDIA DGX Spark Review (2026): One Year On — Strengths, Limits and Verdict
- ReviewNVIDIA DGX Spark Specs: Full GB10 Specifications Explained (and vs RTX Spark)
- ModelBest Models for NVIDIA DGX Spark (2026): What to Run on 128 GB
- Buying guideBest RTX Spark Laptop (2026): How to Choose the Right N1X Laptop
- Buying guideRTX Spark Price Comparison (2026): Every Laptop & Desktop, and What You Actually Get