How much memory does Qwen2.5 32B Instruct need at Q4?

The Q4 figure used here is 22.0 GB. GGUF file size plus an FP16 KV-cache estimate at 8,192 tokens. The cache counts only the layers and heads named in the method on the About page.

What is the cheapest catalog device that runs Qwen2.5 32B Instruct at Q4?

Radeon RX 7900 XTX, list price $999.

How fast is Qwen2.5 32B Instruct on the Mac Studio M3 Ultra 96GB, as a bandwidth estimate?

About 23.5 tok/s at Q8. That is device bandwidth divided by active weight bytes, not a measured run.

Yes. Qwen2.5 32B Instruct fits on the Mac Studio M3 Ultra 96GB at Q8, with 30.2 GB of headroom. Usable memory is 67.2 GB.

Can the Mac Studio M3 Ultra 96GB run Qwen2.5 32B Instruct?

Data updated 16 October 2026. Last reviewed the same day, when this catalog snapshot was written.

Usable memory on this 96 GB unified-memory config is 67.2 GB (70% of the sticker figure). A result under 2 GB of headroom is labeled tight.

Quant ladder

Estimated memory for Qwen2.5 32B Instruct on Mac Studio M3 Ultra 96GB
Quant	Estimate	Headroom	Est. tok/s	Verdict
Q4	22.0 GB	45.2 GB	41.3 tok/s	Yes
Q8	37.0 GB	30.2 GB	23.5 tok/s	Yes
FP16	78.6 GB	-11.4 GB	12.5 tok/s	No

GGUF file size plus an FP16 KV-cache estimate at 8,192 tokens. The cache counts only the layers and heads named in the method on the About page.
Formula estimate: parameter count × bytes per weight × 1.20. No GGUF file size is stored for this quant.
Tokens/sec is a bandwidth roofline: GB/s divided by active weight GB. It is not a measured generation speed.

KV cache at 8K and 32K

FP16 KV-cache estimate: 2.1 GB at 8,192 tokens and 8.6 GB at 32,768 tokens. GQA counts two tensors (K and V) for the stored KV heads and head dimension. config.json also sets a sliding window of 131,072 tokens. The figures above assume every counted layer stores the full context, so they can overstate cache on sliding layers.

Offload

This machine has one memory pool. Weights are not offloaded to a second bank of system RAM.

Ollama

This command pulls the Q4_K_M file from Qwen/Qwen2.5-32B-Instruct-GGUF, the Hub repo whose file size is used above. It is not an official Ollama library tag.

ollama run hf.co/Qwen/Qwen2.5-32B-Instruct-GGUF:Q4_K_M

Other picks

Cheapest catalog device at Q4: Radeon RX 7900 XTX
Closest catalog model that still fits at Q4: Llama 4 Scout.

All hardware for Qwen2.5 32B Instruct · All models for the Mac Studio M3 Ultra 96GB

Questions

How much memory does Qwen2.5 32B Instruct need at Q4?: The Q4 figure used here is 22.0 GB. GGUF file size plus an FP16 KV-cache estimate at 8,192 tokens. The cache counts only the layers and heads named in the method on the About page.
Can the Mac Studio M3 Ultra 96GB run Qwen2.5 32B Instruct?: Yes. Qwen2.5 32B Instruct fits on the Mac Studio M3 Ultra 96GB at Q8, with 30.2 GB of headroom. Usable memory is 67.2 GB.
What is the cheapest catalog device that runs Qwen2.5 32B Instruct at Q4?: Radeon RX 7900 XTX, list price $999.
How fast is Qwen2.5 32B Instruct on the Mac Studio M3 Ultra 96GB, as a bandwidth estimate?: About 23.5 tok/s at Q8. That is device bandwidth divided by active weight bytes, not a measured run.