Writing

AI Infrastructure·18 April 2026

Running Language Models With Limited GPU Resources

Lessons from experimenting with quantisation, QLoRA and smaller language models on consumer hardware.

Useful AI systems should not assume unlimited compute.

Much of my work involves making speech and language models practical on constrained GPU hardware — the kind of environment many African engineering teams actually have.

Practical levers

Techniques that matter in practice:

  • Quantisation — reduce memory footprint with controlled precision trade-offs
  • QLoRA — efficient fine-tuning without full-parameter updates
  • Smaller base models — choose capacity that fits the task
  • Batching and caching — improve throughput for real workloads
  • Local inference first — reduce dependency on unreliable connectivity

Rendering diagram…

Engineering, not magic

Resource-constrained AI is mostly systems engineering:

  • know the latency budget
  • measure VRAM carefully
  • prefer predictable failure modes
  • keep humans in the loop for clinical or high-stakes outputs

The goal is not the largest model. The goal is a reliable system that fits the environment where it must run.