Model → hardware
Qwen3 32B
Plan for 24 GB of dedicated GPU memory at Q4_K_M and a 4,000-token context. 32 GB gives more headroom under the same settings.
Larger general-purpose model. Memory fit is an estimate, not a speed or quality guarantee.
Hardware starting points
Planning estimatesQ4_K_M · 32 GB system RAM · one active request · full GPU residency
| Dedicated GPU memory | At 4,000 tokens | At 16,000 tokens | Example from these guides |
|---|---|---|---|
| 24 GB | Smallest listed tier that fits21.6 GB + 0.72 GB reserve | Exceeds this budget23.3 GB + reserve | RTX 4090 24 GB |
| 32 GB | More headroom21.6 GB + 0.96 GB reserve | Fits the estimate23.3 GB + reserve | RTX 5090 32 GB |
The smallest tier fits the calculation; it is not a verified minimum. Cache precision and runtime buffers can change the fit. “More headroom” means at least 2 GB remains after the estimate and reserve.
Use the publisher’s Instruct / chat template in a current GGUF-capable runtime. The 24 GB tier can become tight when display use, buffers or context grow. CPU offloading needs adequate system RAM and changes performance; it is not included in these full-GPU estimates.
Compare quantisation and context
| File profile | Weight input¹ | GPU estimate at 4K | GPU estimate at 16K |
|---|---|---|---|
| Q4_K_M | 20.20 GB | 21.6 GB | 23.3 GB |
| Q5_K_M | 24.60 GB | 26.0 GB | 27.7 GB |
¹ Catalogue planning weights. These can differ from an exact downloaded file; runtime memory also includes cache and buffers.
Keep the exact file, runtime version, cache type and context limit consistent when comparing. Q5 uses more memory; compare its answers on your own tasks before accepting the trade-off.
Setup and common questions
Which file should I start with?
Use the publisher’s Qwen3-32B-GGUF files at Q4_K_M. Download every shard belonging to the chosen quantisation. The table uses the Finder’s weight input; compare it with the exact files and your runtime’s reported memory.
Should thinking mode be on?
This checkpoint supports thinking and non-thinking modes. Choose one explicitly and keep it fixed when comparing answers or timings. Thinking consumes output tokens and time before the final answer.
Can I use the advertised long context?
Qwen lists 32,768 native tokens and up to 131,072 with YaRN. Extended context needs runtime support and extra cache memory. The short-context fit here does not verify YaRN or the full window. The model uses Apache 2.0.
How are these recommendations chosen?
We compare dedicated GPU memory tiers using the main Finder’s memory calculation at the stated quantisation and context. These tiers describe memory fit, not speed or output quality.
GPU memory includes weights, context cache and a 0.9 GB runtime allowance, plus a 3% reserve (0.15–1.5 GB). Cache format and runtime buffers can change the fit. See the calculation assumptions.
Sources and related guides
Specification sources checked 3 October 2026.
These comparisons are planning estimates. No reviewed or independently reproduced measurements are available yet. About test evidence.