EXO Labs runs Llama 2 on 1997 Pentium II with 128 MB

EXO Labs loaded a slimmed variant of Llama 2 onto a 1997 Pentium II desktop with 128 MB of RAM and ran inference using BitNet’s ternary-weight method, which reduces neural network parameters to three values: -1, 0 and 1.

The team combined ternary weights with quantization, pruning and optimized data layout to cut memory and compute needs. Replacing 16- or 32-bit floating-point weights with three-value representations reduced storage and simplified arithmetic. On the single-core CPU the model generated coherent text very slowly, producing output word by word.

EXO Labs described the exercise as a software-first efficiency experiment rather than an attempt to replace modern accelerators. The demonstration used a pared-down model focused on inference; it did not include large-scale training and did not match the throughput of GPU-based deployments.

The technical approach reduced the amount of state the processor had to handle by lowering precision and removing unneeded connections. Optimized memory layout lowered bandwidth demands, allowing the vintage machine to complete inference instead of failing due to memory constraints.

The team noted that similar quantization and pruning techniques can be applied to models intended for modern low-power hardware, which could enable on-device inference on midrange laptops, edge gateways or compact microservers and reduce reliance on cloud GPUs for some workloads. EXO Labs did not claim feature parity or performance comparable to full-size models running on dedicated accelerators.

The demonstration relates to discussions about AI energy use by showing how software techniques can reduce hardware demands for basic inference on older silicon. EXO Labs published the methods used — ternary weights, pruning and layout optimizations — and added that further work is needed to adapt the techniques for practical, everyday applications.