Table of Contents
Qwen 2.5 14B FP16 is 28 GB — does not fit a 16 GB card. AWQ-INT4 (~9 GB) does fit. Quality cost is ~1-2% on standard benchmarks, but the cost-per-token saving is real.
TL;DR
Qwen 2.5 14B at AWQ-INT4 fits the RTX 5060 Ti with ~5 GB headroom for KV cache. ~440 tok/s aggregate, ~52 tok/s single-stream. £0.24/1M tokens. For FP16 quality, step up to a 5090 32 GB.
VRAM fit
| Precision | Weights VRAM | Available KV pool | Concurrent users |
|---|---|---|---|
| FP16 | 28 GB | Does not fit | n/a |
| FP8 | 14 GB | ~1 GB | ~3 users |
| AWQ-INT4 | ~9 GB | ~5 GB | ~12 users |
Benchmark results
| Metric | 5060 Ti AWQ-INT4 | 5090 FP8 | 5090 FP16 |
|---|---|---|---|
| Aggregate tok/s | 440 | 1,150 | 720 |
| Single-stream tok/s | 52 | 80 | 58 |
| Median TTFT | 320 ms | 180 ms | 220 ms |
| Cost per 1M tokens | £0.24 | £0.18 | £0.32 |
Verdict
5060 Ti runs Qwen 2.5 14B at AWQ-INT4 — workable for low-concurrency deployments. For production at meaningful scale, 5090 + FP8 is dramatically better and roughly the same cost-per-token.
Bottom line
For 14B-class models the 5090 32 GB is the right home. The 5060 Ti is a budget fallback. See Qwen 14B deployment.
