Compare
Quality tied — speed and VRAM decide
- 2026-09-08
- local LLM
- RTX 5080
- coding
Four identical coding tasks on the Rumble desktop (RTX 5080 16GB).
What happened
Tasks: merge intervals, fix-from-traceback, multi-file wire, mini SWE.
- Gemma 4 12B (agent-hi): 4/4 · ~70 tok/s · ~10.4GB VRAM
- Qwen 27B Unsloth IQ4_XS: 4/4 · ~28 tok/s · ~15.6GB VRAM (tight)
Same quality score. Different daily-driver feel.
Takeaway
On 16GB, keep the snappy 12B as the interactive default. Keep the bigger IQ4 around for more brain / less headroom. House aliases now expose both.
Numbers from a single measured stack (Ollama, thinking off). Not a claim about other GPUs.
Sister site: toklanes.com — measured tok/s review-bench.