What hardware do you need to run Qwen3 14B?
Qwen3 14B at Q4_K_M is 9.0 GB of weights, with a context ceiling of 40k tokens. The sweet spot for a 24 GB machine. Everyday coding help is usable; long agent loops drift.
How good is it, really?
On the Artificial Analysis Intelligence Index v4.3 it scores 6 (reasoning mode; 7 without), which puts it in the Below every hosted tier band. Fine for simple, well-specified tasks. Noticeably less capable than anything the big labs sell today. Score source. See the whole table.
- Summarising — good
- Translation — good
- Everyday coding — usable
- Reasoning & maths — usable
- Agentic work — don’t
What it costs either way
Renting the same model costs $0.1 per million input tokens and $0.22 per million output (OpenRouter, cheapest active endpoint, checked 2026-09-03). Buying a machine only beats that if you use it hard enough, for long enough, that the hardware price divides down below the rental bill.
Machines that run it
| Machine | Price | Speed at 32k | Pay-back | |
|---|---|---|---|---|
| Mac mini M6, 24GB | $1,099 | 8.9 tok/s estimated | Pays back in 70 years | Run the numbers |
| Strix Halo Framework Desktop, 32GB | $1,269 | 13 tok/s estimated | Pays back in 89 years | Run the numbers |
| MacBook Pro M5 (14-inch), 32GB | $2,399 | 8 tok/s estimated | Pays back in 157 years | Run the numbers |
| Mac Studio M5 Max, 36GB | $2,499 | 24 tok/s estimated | Pays back in 153 years | Run the numbers |
| DGX Spark GB10 Grace Blackwell, 128GB | $4,699 | 14 tok/s estimated | Pays back in 337 years | Run the numbers |
One machine per family, cheapest first. Speeds are measured where a public benchmark exists and estimated from memory bandwidth otherwise; the calculator says which for any configuration.
The specifics
- Parameters
- 14.8B
- Quantisation
- Q4_K_M
- Weights on disk
- 9.0 GB
- KV cache
- 5.4 GB at 32k context
- Maximum context
- 40k tokens (32k natively; 128k with YaRN)
- Licence
- Apache 2.0
- Sources
- source 1, source 2