Can I Run Codestral on Framework Desktop (Ryzen AI Max+ 395, 128 GB)?

Superseded model. Codestral has been superseded by Devstral. This page is kept for reference; the newer family is a better starting point. View Devstral →

Written by Jakub Rusinowski · Last updated May 29, 2024

Yes — comfortably

Yes, comfortably — Codestral 22B at Q8_0 needs about 26.3 GB of the 128 GB usable on Framework Desktop (Ryzen AI Max+ 395, 128 GB), leaving ~101.7 GB spare and running at ~7.7 tok/s (estimated), with room for about 32,768 tokens of context.

Confidence: high · Recommended quantization: Q8_0 · Estimated speed: ~7.7 tok/s

Framework Desktop (Ryzen AI Max+ 395, 128 GB) — what it gives a model

Usable memory for models	128 GB
Memory bandwidth	256 GB/s
Form factor	Mini PC
Operating system	Windows or Linux
Memory upgradeable	No — soldered

Codestral on Framework Desktop (Ryzen AI Max+ 395, 128 GB): memory by quantization

Quant	Memory needed	Fits 128 GB?	Max context	Est. speed	Download
F16	47.1 GB	✓ Yes	32K	~4.2 tok/s	44.4 GB
Q8_0	26.3 GB	✓ Yes	32K	~7.7 tok/s	23.6 GB
Q6_K	20.9 GB	✓ Yes	32K	~9.8 tok/s	18.2 GB
Q5_K_M	18.4 GB	✓ Yes	32K	~11.2 tok/s	15.7 GB
Q4_K_M	16.1 GB	✓ Yes	32K	~13 tok/s	13.4 GB
Q3_K_M	12.1 GB	✓ Yes	32K	~17.7 tok/s	9.5 GB
Q2_K	10 GB	✓ Yes	32K	~22.1 tok/s	7.3 GB

What to watch out for

Memory on this machine is not upgradeable, so the configuration you buy is the ceiling for every model you will ever run on it.

Framework Desktop 128 GB limitations

Memory is soldered LPDDR5X — unusually for Framework, this is the one component that cannot be upgraded.
ROCm/Vulkan support for Strix Halo is younger than CUDA; check your runtime supports it before buying.

Recommended setup

Ollama or llama.cpp (CUDA/ROCm) — vLLM if you need concurrent serving

How these numbers are calculated

Memory need = quantized weights + KV cache at 8,192 tokens (f16) + 0.8 GB runtime overhead.
128 GB unified memory at 256 GB/s, shared between CPU and GPU.
Throughput is a memory-bandwidth roofline estimate, not a measurement. It is labelled "estimated" everywhere it appears.
KV cache is computed from this model's published attention configuration.

FAQ