Only with CPU offload
Qwen2.5-Coder 32B at Q4_K_M needs 21.0 GB but only 7.0 GB is addressable, so about 76% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 2.5 tokens per second — usable for batch work, painful for chat.
24% on GPU
8K context
Q4_K_M · 18.4 GB
Apache 2.0
Released Nov 2024
The first local code model that felt competitive with hosted assistants. Qwen3.8 27B is smaller and far ahead on agentic work.
What hardware do I need for Qwen2.5-Coder 32B? →
Fits instead: Qwen2.5-Coder 7B (5.3 GB)
The VRAM budget
weights 18.4 GB
Weights 18.4 GB
KV cache @ 8K 2.00 GB
Runtime overhead 0.6 GB
Over budget 14.0 GB past 7.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 32.5 GB | 35.1 GB | — | ~1.3 | −0.1% ppl | 28.1 GB over |
| Q6_K | 25.0 GB | 27.6 GB | — | ~1.7 | −0.4% ppl | 20.6 GB over |
| Q5_K_M | 21.7 GB | 24.3 GB | — | ~2.1 | −0.8% ppl | 17.3 GB over |
| Q4_K_M | 18.4 GB | 21.0 GB | — | ~2.5 | −1.9% ppl | 14.0 GB over |
| Q3_K_M | 14.9 GB | 17.5 GB | — | ~3.3 | −5.4% ppl | 10.5 GB over |
| Q2_K | 12.8 GB | 15.4 GB | — | ~4.1 | −15% ppl | 8.4 GB over |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it.
How to run it
$ llama-server \
-hf Qwen/Qwen2.5-Coder-32B-Instruct:Q4_K_M \
-c 8192 -ngl 15
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 18.4 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.
03If it falls back to CPU silently, drop the context first, then step down a quantisation.