A Reddit user upgraded a second-hand laptop to 96GB of memory and spent about £3,100 in total, pushing a quantized Qwen model to nearly 49 tokens per second in generation and read speed. The result suggests that the barrier to running large models locally is shifting from dedicated hardware toward modifiable consumer devices.
What this is
This is not a new product from a vendor but a personal DIY hardware project: an MSI Vector 16 with an RTX 5080 and 32GB of memory, fitted with second-hand 96GB laptop memory. The device also retains two Thunderbolt 5 ports for connecting an external GPU. The author also tested the Q3-quantized version of GLM 5.3 Flash, achieving roughly 25 tokens per second.
The key issue is not “how large a model can a laptop run,” but how much VRAM and system memory determine whether the model can be fully resident. Quantization compresses model parameters into lower precision, reducing hardware requirements in exchange for acceptable speed.
Industry view
This case is worth watching because the cost structure of local AI is starting to loosen. High-end desktop mini PCs can cost around £3,500, while so-called AI laptops with a stated 128GB of memory can run about £3,200. The modified system is actually cheaper while preserving portability. For people who need to process sensitive material or reduce cloud-inference costs, this is a viable path.
But the objections are equally clear. This is one user's single test, not evidence of stability, thermal performance, noise levels, or long-term compatibility. Second-hand 96GB memory costs about £800, and accessory prices and provenance need verification. The actual bandwidth of the Thunderbolt interface, external power delivery, software support, and model quality could all reduce the headline performance. More importantly, the article notes that the model’s “role-playing ability” was inadequate, showing that merely running a model is not the same as using it well.
Impact on regular people
For enterprise IT: Local models can supplement cloud services for specific workloads where data cannot leave the organization. Before procurement, teams should test the model, drivers, and concurrent workloads rather than relying on tokens per second alone.
For individual professionals: High-capacity laptops could handle tasks such as document summarization and code assistance. However, upgrade costs, battery life, and thermal performance remain practical constraints; portability and performance usually cannot both be maximized.
For the consumer market: Memory capacity and expandability may matter more than simply advertising a device as an “AI laptop.” Still, the current approach remains niche and does not prove that expensive AI laptops are broadly poor value.