---
title: Is 128GB Enough or Do You Need 192GB? The Boundary Line of Local AI Unified Memory by Capacity
date: 2026-09-23
time: 23:59
model: admin
category: knowhow
summary: 70B runs comfortably on 128GB while 150GB-class monsters need 192GB, and this lays out the boundary lines of unified memory selection, including why capacity does not guarantee speed
tags: unified-memory, 128GB, 192GB, Strix-Halo, Mac-Studio, DGX-Spark, local-LLM, VRAM, KV-cache, bandwidth
---

To start with the conclusion: for most people 128GB is enough, and 192GB is a choice only for those who can clearly name the ultra-large model they want to run and the reason why. And more capacity does not make it faster. Look at the boundary lines below before opening your wallet.

## 1. 128GB vs 192GB: What Can You Do?

| Item | 128GB environment | 192GB environment |
|---|---|---|
| Main target models | Llama 3.1 70B Q4, Qwen 2.5 72B Q4 (40-45GB) | Qwen 2.5 72B Q8 (about 77-133GB), DeepSeek-V4 Flash Q4 (about 99-155GB) |
| Ultra-large models | 235B MoE is barely possible with low quantization (Q3 or below) | 235B Q4 with room to spare, up to Llama 405B Q2 (about 177GB) |
| Headroom | Loading a large model leaves limited context (conversation history) | Large model + long conversation + background AI services simultaneously |
| Representative machines | Strix Halo mini-PC, DGX Spark | Mac Studio M2 Ultra 192GB, Framework 192GB build (launch announced) |
| Recommended for | Value-focused buyers, first serious local AI build | Those who want to run 150GB-class models as-is without compromise |

To prove it with numbers, 70B Q4 is 40-45GB, so even after subtracting system overhead from 128GB there is ample context left. For coding assistance or everyday text generation, 128GB is fast enough and practical. On the other hand, running Qwen 2.5 72B at Q8 (low compression, high quality) or loading a DeepSeek-V4 Flash 284B-class model crosses the 128GB wall, so 192GB is required.

## 2. Actual Memory Footprint by Quantization: How Many GB Is My Model?

The size a model occupies when loaded into memory is the parameter count multiplied by the quantization bits. The table below is the starting point for every judgment.

| Model | BF16/FP16 | Q8_0 | Q4_K_M | Q3 or below |
|---|---|---|---|---|
| 7-8B (Llama 3.1 8B, etc.) | ~16GB | ~8.5GB | ~5GB | ~3.5GB |
| 14B (Phi-4, Qwen 14B) | ~28GB | ~15GB | ~9GB | ~6GB |
| 27-32B (Qwen 27B, QWQ 32B) | ~60GB | ~33GB | ~19GB | ~13GB |
| 70-72B (Llama 70B, Qwen 72B) | ~145-160GB | ~77GB | ~43GB | ~30GB |
| 120B MoE (GPT-OSS 120B) | ~240GB | ~130GB | ~70GB | ~50GB |
| 235B MoE (Qwen3 235B) | ~470GB | ~260GB | ~142GB | ~100GB |
| 284B MoE (DeepSeek V4 Flash) | ~560GB | ~300GB | ~155GB | ~110GB |
| 405B (Llama 3.1 405B) | ~810GB | ~430GB | ~240GB | ~178GB (Q2) |

To this you must add the system reservation (about 8GB) and the KV cache (several GB to tens of GB during long conversations) to get the real usable capacity. If a 128GB machine has about 120GB of real usable memory, 70B Q4 (43GB) has room to spare but 235B Q4 (142GB) physically cannot load. If a 192GB machine has about 184GB usable, 235B Q4 fits with room left over and 405B Q2 (178GB) just barely fits. This is the physical boundary line between the two capacities.

## 3. KV Cache: The Real Culprit That Blows Up When You Make Conversations Longer Than the Model

Model weights are a fixed cost, but the KV cache is a variable cost that grows as conversations get longer. Stack 100K tokens of conversation on a 70B model and the KV cache alone adds tens of GB. Running 70B Q4 on 128GB while analyzing a long repository easily exceeds 100GB with the model (43GB) + KV cache (30GB or more) + system, and from that point context limits kick in. With 192GB, even under the same conditions more than 80GB remains, so you can run a background translation agent or an embedding server alongside. If long-context work is your main job, the weight of the capacity choice changes.

## 4. The Truth About Speed: Token Speed Is Computed From Bandwidth

The formula for local inference speed is simple. Tokens per second (TPS) is proportional to memory bandwidth divided by model size.

| Machine | Bandwidth | 70B Q4 estimated speed | 235B Q4 estimated speed |
|---|---|---|---|
| Strix Halo 128GB (256GB/s) | 256GB/s | ~8-12 tok/s | ~3-5 tok/s (low quantization) |
| DGX Spark (273GB/s) | 273GB/s | ~10-13 tok/s | ~4-6 tok/s |
| Mac Studio M2 Ultra (800GB/s) | 800GB/s | ~25-35 tok/s | ~13-20 tok/s |

Real measurements back this up. Qwen3 235B was reported at 11 tok/s on a GMKtec EVO-X2 128GB, and DeepSeek V4 Flash 284B at 52 tok/s on a Mac Studio M2 Ultra 192GB. Even for the same model, a 3x bandwidth difference shows up directly as a 3x speed difference.

That is why there is a trap in the announced 192GB-class Strix Halo model. Capacity grows by 50% but bandwidth improves by only 7%. Many people will buy it looking only at the 192GB number and then be disappointed to find the speed unchanged. If you want speed, look at the bandwidth number, not the capacity.

## 5. Representative Machines and Real Prices (Second Half of 2026)

First, note that due to the DRAM shortage, prices have nearly doubled compared to late 2025. The following are actual street prices.

| Machine | Memory | Bandwidth | Price range | Notes |
|---|---|---|---|---|
| GMKtec EVO-X2 | 128GB LPDDR5X | 256GB/s | List ~$2,000 but scarce, street ~$3,400 | Advertised at Qwen3 235B 11 tok/s |
| Framework Desktop | 128GB | 256GB/s | ~$3,449, often sold out | 192GB build announced, easy to disassemble and repair |
| Minisforum MS-S1 Max | 128GB | 256GB/s | ~$3,639, in stock | Dual 10GbE, expansion slot, top-rated reviews |
| NVIDIA DGX Spark | 128GB | 273GB/s | Raised from $3,999 to $4,699 | CUDA ecosystem, NVFP4 support, 256GB when two are linked |
| Mac Studio M2 Ultra | 192GB | 800GB/s | High price before discontinuation, secondhand market | Unrivaled #1 in speed, quiet |

The Strix Halo camp uses the same chip, so speeds are roughly similar at around 10 tok/s for Qwen3.8 27B. The criteria for choosing are not price but stock, ports, and support. The box you can actually buy is the answer. With DGX Spark, linking two units with a ConnectX cable lets you use them as a single 256GB pool, providing an expansion card that can reach up to 405B-class models.

The only current real purchase option for 192GB is, in practice, the Mac Studio M2 Ultra 192GB. Thanks to its 800GB/s bandwidth, there are real measurements of DeepSeek V4 Flash 284B running at 52 tok/s, and Qwen3 235B Q4 is handled comfortably at 13-20 tok/s.

## 6. Recommendation Matrix by Use Case

| Use case | Recommended capacity | Reason |
|---|---|---|
| Coding assistance, everyday chat (8-32B) | 64GB is enough, 128GB is overkill | 32B Q4 is 19GB, so 64GB has room |
| Always-on 70B-class agent + RAG | 128GB | Model 43GB + embeddings + long context fit just right |
| Whole-repository analysis, 100K+ long context | 128GB minimum, 192GB recommended | The KV cache eats tens of GB |
| Always-on 235B MoE high quality (Q4) | 192GB | 128GB requires compromising quality with low quantization |
| 405B-class sampling, multiple models at once | 192GB or more | Even at Q2 quantization it is 178GB, so 192GB is the bottom line |
| Direct training including fine-tuning | A separate CUDA GPU is required | Unified memory is for inference; training lacks bandwidth |

## 7. The Hidden Costs of Electricity and Noise

If you run it 24/7, the electricity bill comes every month. The Strix Halo mini-PC draws 130-150W under load, the Mac Studio around 100W, and the DGX Spark 240W (around 100W in real use). Compared with a desktop RTX 4090 system (600W or more), the monthly electricity bill drops to a third or less. Noise is also at library levels for mini-PCs and Macs, so you can sleep with one in your room, and even DGX-class fan noise is not loud enough to prevent placing it in the living room. If you are aiming for a bedroom server, low power is the answer.

## 8. Three Essential Warnings Before Buying

First, more capacity does not make it faster. The formula in section 4 is everything. Do not be dazzled by the 192GB sticker; check the bandwidth number.

Second, memory cannot be added later. Strix Halo mini-PCs and Macs use onboard LPDDR5X, so the time of purchase is your last chance. Regret later comes too late, so settle your use case in your first budget. In particular, you must decide whether to waste half a year waiting for the Framework 192GB build or to start now with 128GB.

Third, beware of secondhand and scarcity premiums. During a DRAM shock, sold-out boxes command a markup. If it exceeds 1.5x the list price, a box one tier lower (64GB) plus an API combination is cheaper. If you are not in a hurry, waiting until DRAM prices stabilize is also a strategy.

## 9. Final Buying Guide

Most people should buy 128GB. The 70B-class main model runs comfortably in the 40GB range, and the budget is kept to about half. The extra cost of 192GB (about $1,500 or more) is justified only when you can clearly name the model you want to run (Qwen 2.5 72B Q8, DeepSeek-V4 Flash, etc.) and the benefit you gain from it (long context, concurrent services).

If you pin down your concrete use cases (programming, data analysis, writing), the models you want to run, and budget criteria, machine selection will be done in 10 minutes.