31 Aug 2026
llama3.2:3b on RTX 3060 Laptop
Intel Core i5-10500H, RTX 3060 Laptop 6GB VRAM · 62.5GB RAM · Ubuntu 24.04.4 LTS
- Model
- llama3.2:3b
- Quantization
- Q4 (default Ollama pull)
- Tokens / sec
- 114.23 tok/s
- RAM used
- 2.85 GB
- Cold start
- 2.244 s
- Notes
- Tested twice: once with a full dev environment running in the background (Docker, Keycloak, Django, a local Kubernetes cluster), once on a clean machine. Tokens/sec barely changed (113.0 vs 114.23) since generation runs entirely on the GPU — but cold start roughly doubled under load (4.93s vs 2.24s), likely from disk I/O contention. VRAM usage was identical either way at 2.85GB, well under this GPU's 6GB budget.
Real numbers, tested twice — once busy, once clean.
We ran llama3.2:3b on a mid-range gaming laptop (Intel i5-10500H, RTX 3060 Laptop, 6GB VRAM) under two conditions: once with a full development environment running in the background, once on a clean machine with nothing else open.
Generation speed barely moved — 113.0 tokens/sec busy, 114.23 tokens/sec clean — because this model runs entirely on the GPU, largely independent of what the CPU is doing elsewhere. Cold start roughly doubled under load, from 2.24s clean to 4.93s with Docker and other processes competing for disk access — a real, measurable cost of a busy machine, not something we’re hiding.
VRAM usage held steady at 2.85GB either condition — comfortably under this GPU’s 6GB ceiling, with headroom for a larger model on the same hardware tier.