What This Is
A Reddit user plans to upgrade his existing Z690 system from 64GB to 128GB of memory so he can run larger models such as Qwen3.8 Flash locally. His current setup can already handle three sessions in parallel at an aggregate decoding speed of roughly 100 tokens per second, but the models and caches together consume about 55GB. He wants the additional memory for more caching, longer context windows, and experiments with larger models.
The value of this personal hardware discussion is not whether he should buy a 4×32GB memory kit. It is that the case reveals a new bottleneck for local large language models—open-source models run on your own computer: whether a model fits entirely in video memory or system memory is now determining the model sizes and usage patterns that are practical.
Industry View
We are seeing RAM capacity and VRAM capacity increasingly set the ceiling for local models. With 128GB of total memory, users can retain more cached data and gain more room to operate when a model is split between VRAM and system memory. That does not, however, mean inference speed will rise proportionately. Memory bandwidth, CPU data-transfer performance, and model quantization—compressing a model to lower precision to reduce resource requirements—can still be the practical limits.
The risks are just as clear: this configuration could cost more than €500, and even at identical speeds, four memory modules may introduce compatibility, stability, and performance compromises depending on the CPU's memory controller, motherboard layout, or cooling conditions. More importantly, if an application relies primarily on a cloud API, the utilization rate of large amounts of personally purchased memory may be low. If a task requires extremely fast responses, added capacity cannot compensate for insufficient GPU compute.
Impact on Regular People
For enterprise IT: When evaluating local models, teams should put memory, VRAM, bandwidth, and long-term utilization on the same cost table. Large-memory workstations make the most sense when data cannot leave the enterprise, compliance requirements are explicit, and usage is consistently high enough to justify the investment.
For individual professionals: Local deployment offers more controllable document processing and coding assistance, but 128GB of memory remains an enthusiast-class configuration. Most knowledge work does not require building a private model stack; trying mature products first is usually more economical.
For the consumer market: As large models move from the cloud onto personal devices, demand for high-capacity memory and compact workstations will continue to grow. What consumers will actually pay for, however, may be a complete, plug-and-play experience rather than a set of hardware components that requires repeated tuning.