48GB VRAM

Radeon PRO W7900 48G — what LLMs can it run?

94 of 110 indexed models fit comfortably in 48GB at 4K context, each at the highest-quality quant that still leaves headroom.

The short answer for Radeon PRO W7900 48G

48 GB, 864 GB/s GDDR6 ECC. 94 of 110 models in this index fit comfortably at 4K context; 95 load at all.

Biggest that fits: GLM-4.5-Air — 37.2 GB at Q2_K — 10.8 GB spare at 4K, so longer context comes out of a thin margin.
Room to grow: Nemotron 3.5 Bolt 74B-A9B — 23.1 GB at IQ2_XXS — under 60% of the card, which leaves 25.0 GB for a long context window or a second process.
Speed ceiling: Llama 3.1 8B Instruct at Q4_K_M reads 4.6 GB of weights per token, so 864 GB/s puts a hard ceiling near 187 tok/s. That is arithmetic on two published numbers, not a benchmark — real throughput lands below it. No run on this card has been measured here, so there is nothing to compare the ceiling against.
The wall you will hit: Backend coverage, not capacity. At 864 GB/s the memory system is competitive with NVIDIA cards of the same size; what varies is whether your runtime has a working ROCm build for your kernel and card. llama.cpp and vLLM both ship official ROCm support — this is a setup question, not a hardware ceiling.

70B+ · 7

Radeon PRO W7900 48G — what LLMs can it run? — 70B+
Model	Quant	Est. VRAM	Headroom	tok/s
GLM-4.5-Air106B MoE	Q2_K	37.2 GB	+10.8 GB	≤231Ceiling
Nemotron 3.5 Bolt 74B-A9B74B-A9B	IQ2_XXS	23.05 GB	+25 GB	≤340Ceiling
Qwen2.5 72B Instruct72B	AWQ INT4	39.36 GB	+8.6 GB	≤25Ceiling
Llama 3.1 70B Instruct70B	AWQ INT4	38.26 GB	+9.7 GB	≤26Ceiling
Llama 3.3 70B Instruct70B	AWQ INT4	38.26 GB	+9.7 GB	≤26Ceiling
DeepSeek-R1-Distill-Llama-70B70B	AWQ INT4	38.26 GB	+9.7 GB	≤26Ceiling
Jamba 1.5 Mini52B-A12B	Q4_K_M	33.01 GB	+15 GB	≤125Ceiling

32B · 30

Radeon PRO W7900 48G — what LLMs can it run? — 32B
Model	Quant	Est. VRAM	Headroom	tok/s
Kimi Linear 48B-A3B Instruct48B-A3B	Q4_K_M	30.44 GB	+17.6 GB	≤500Ceiling
Mixtral 8x7B Instruct47B MoE	Q4_K_M	30.13 GB	+17.9 GB	≤116Ceiling
Seed-OSS 36B Instruct36B	Q5_K_M	27.81 GB	+20.2 GB	≤36Ceiling
Qwen3.6 35B-A3B35B-A3B	Q8_0	39.6 GB	+8.4 GB	≤285Ceiling
Command R 35B35B	Q4_K_M	22.86 GB	+25.1 GB	≤43Ceiling
Yi 1.5 34B Chat34B	Q4_K_M	22.82 GB	+25.2 GB	≤44Ceiling
Qwen3 32B Instruct32B	Q4_K_M	21.85 GB	+26.2 GB	23–31Derived
Qwen2.5 32B Instruct32B	Q4_K_M	21.69 GB	+26.3 GB	23–31Derived
Qwen2.5-Coder 32B Instruct32B	Q4_K_M	21.69 GB	+26.3 GB	23–31Derived
DeepSeek-R1-Distill-Qwen-32B32B	Q4_K_M	21.69 GB	+26.3 GB	23–31Derived
Nemotron 3.5 Lightning 30B-A3B30B-A3B	Q8_0	35.14 GB	+12.9 GB	≤285Ceiling
Gemma 4 31B IT31B	Q8_0	35.72 GB	+12.3 GB	14–19Derived
Qwen3 30B-A3B Instruct30B-A3B	Q5_K_M	23.04 GB	+25 GB	≤427Ceiling
Qwen3-Coder 30B-A3B Instruct30B-A3B	Q5_K_M	23.04 GB	+25 GB	≤427Ceiling
Qwen3-VL 30B-A3B Instruct30B-A3B	Q8_0	34.28 GB	+13.7 GB	≤285Ceiling
GLM-4.7-Flash30B-A3B	Q8_0	33.47 GB	+14.5 GB	≤285Ceiling
Granite 4.2 30B30B	Q8_0	33.61 GB	+14.4 GB	15–20Derived
Muse Glimmer 30B30B	Q8_0	31.08 GB	+16.9 GB	16–21Derived
Qwen3.8 27B27B	Q4_K_M	17.47 GB	+30.5 GB	28–37Derived
Qwen3.6 27B27B	Q8_0	30.71 GB	+17.3 GB	16–21Derived
Gemma 3 27B IT27B	Q4_K_M	18.37 GB	+29.6 GB	28–37Derived
Gemma 2 27B Instruct27B	Q5_K_M	21.76 GB	+26.2 GB	24–31Derived
Gemma 4 26B-A4B IT26B-A4B	Q8_0	28.42 GB	+19.6 GB	≤214Ceiling
Mistral Small 24B Instruct24B	Q4_K_M	15.89 GB	+32.1 GB	32–42Derived
LFM2.5 24B-A2B24B-A2B	Q8_0	26.73 GB	+21.3 GB	≤349Ceiling
Devstral Small 1.1 24B24B	Q6_K	20.91 GB	+27.1 GB	24–31Derived
Magistral Small 1.2 24B24B	Q6_K	20.91 GB	+27.1 GB	24–31Derived
Codestral 22B22B	Q4_K_M	15.03 GB	+33 GB	34–45Derived
ERNIE 4.5 21B-A3B21B-A3B	Q8_0	24.44 GB	+23.6 GB	≤285Ceiling
GPT-OSS 20B21B MoE	MXFP4	12.76 GB	+35.2 GB	≤437Ceiling

14B · 18

Radeon PRO W7900 48G — what LLMs can it run? — 14B
Model	Quant	Est. VRAM	Headroom	tok/s
InternLM2 20B Chat20B	Q5_K_M	15.59 GB	+32.4 GB	32–43Derived
DeepSeek-Coder-V2-Lite Instruct16B-A2.4B	Q8_0	17.56 GB	+30.4 GB	≤357Ceiling
DeepSeek-V2-Lite Chat16B-A2.4B	Q4_K_M	10.08 GB	+37.9 GB	≤625Ceiling
StarCoder2 15B15B	Q4_K_M	10.19 GB	+37.8 GB	50–67Derived
Qwen3 14B Instruct14B	Q5_K_M	11.65 GB	+36.4 GB	44–58Derived
Qwen2.5 14B Instruct14B	Q5_K_M	11.73 GB	+36.3 GB	44–58Derived
DeepSeek-R1-Distill-Qwen-14B14B	Q4_K_M	10.14 GB	+37.9 GB	51–68Derived
Phi-4 14B14B	Q5_K_M	11.77 GB	+36.2 GB	44–58Derived
Phi-3 Medium 14B Instruct14B	Q6_K	12.86 GB	+35.1 GB	40–53Derived
Mistral Nemo 12B Instruct12B	Q6_K	11.14 GB	+36.9 GB	46–61Derived
Gemma 3 12B IT12B	Q5_K_M	9.84 GB	+38.2 GB	53–70Derived
Mellum2.1 12B-A2.5B Thinking12B-A2.5B	Q8_0	13.6 GB	+34.4 GB	≤343Ceiling
Stable LM 2 12B Chat12B	Q4_K_M	8.35 GB	+39.7 GB	63–83Derived
Llama 3.2 11B Vision Instruct11B	Q8_0	12.9 GB	+35.1 GB	39–52Derived
Solar 10.7B Instruct11B	Q4_K_M	7.6 GB	+40.4 GB	71–94Derived
Falcon 3 10B Instruct10B	Q4_K_M	7.28 GB	+40.7 GB	73–96Derived
Gemma 2 9B Instruct9B	Q8_0	11.7 GB	+36.3 GB	47–62Derived
Qwen3.5 9B9B	Q8_0	10.37 GB	+37.6 GB	47–62Derived

7B · 25

Radeon PRO W7900 48G — what LLMs can it run? — 7B
Model	Quant	Est. VRAM	Headroom	tok/s
GLM-4-9B-Chat9B	Q8_0	10.16 GB	+37.8 GB	48–64Derived
Qwen3-VL 8B Instruct8B	Q8_0	10.39 GB	+37.6 GB	49–65Derived
Granite 4.2 8B8B	Q8_0	10.45 GB	+37.6 GB	49–65Derived
Ministral 3 8B Instruct8B	Q8_0	10.02 GB	+38 GB	51–67Derived
Qwen2-VL 7B Instruct7B	Q4_K_M	5.49 GB	+42.5 GB	91–121Derived
Granite 3.1 8B Instruct8B	Q4_K_M	5.88 GB	+42.1 GB	92–122Derived
Qwen3 8B Instruct8B	Q6_K	7.64 GB	+40.4 GB	68–90Derived
Llama 3.1 8B Instruct8B	Q8_0	9.47 GB	+38.5 GB	54–71Derived
Nous Hermes 3 Llama 3.1 8B8B	Q4_K_M	5.64 GB	+42.4 GB	94–125Derived
Aya 23 8B8B	Q4_K_M	5.64 GB	+42.4 GB	94–125Derived
OpenChat 3.6 8B8B	Q4_K_M	5.64 GB	+42.4 GB	94–125Derived
DeepSeek-R1-Distill-Llama-8B8B	Q5_K_M	6.51 GB	+41.5 GB	80–106Derived
InternLM2 7B Chat7B	Q4_K_M	5.45 GB	+42.6 GB	98–129Derived
Qwen2.5 7B Instruct7B	Q6_K	6.77 GB	+41.2 GB	73–97Derived
Qwen2.5-Coder 7B Instruct7B	Q4_K_M	5.07 GB	+42.9 GB	99–131Derived
DeepSeek-R1-Distill-Qwen-7B7B	Q4_K_M	5.07 GB	+42.9 GB	99–131Derived
Gemma 4 E4B ITE4B	Q8_0	8.46 GB	+39.5 GB	57–76Derived
OLMo 2 7B Instruct7B	Q8_0	10.3 GB	+37.7 GB	59–78Derived
OLMo 3 7B Instruct7B	Q8_0	10.3 GB	+37.7 GB	59–78Derived
WizardLM-2 7B7B	Q4_K_M	5.14 GB	+42.9 GB	105–138Derived
Mistral 7B Instruct v0.37B	Q6_K	6.75 GB	+41.3 GB	77–102Derived
Zephyr 7B Beta7B	Q6_K	6.75 GB	+41.3 GB	77–102Derived
Gemma 4 E2B ITE2B	Q8_0	5.2 GB	+42.8 GB	≤184Ceiling
Qwen3.5 4B4B	Q8_0	4.97 GB	+43 GB	≤197Ceiling
Gemma 3 4B IT4B	Q8_0	5.05 GB	+43 GB	≤199Ceiling

≤3B · 14

Radeon PRO W7900 48G — what LLMs can it run? — ≤3B
Model	Quant	Est. VRAM	Headroom	tok/s
Qwen3 4B Instruct4B	Q6_K	4.05 GB	+44 GB	≤277Ceiling
Phi-4 Mini Instruct3.8B	Q8_0	4.81 GB	+43.2 GB	≤223Ceiling
Phi-3.5 Mini Instruct3.8B	Q8_0	5.89 GB	+42.1 GB	≤224Ceiling
Granite 4.2 3B3B	Q8_0	4.41 GB	+43.6 GB	≤234Ceiling
Llama 3.2 3B Instruct3B	Q8_0	4.05 GB	+44 GB	≤267Ceiling
Qwen2.5 3B Instruct3B	Q8_0	3.59 GB	+44.4 GB	≤277Ceiling
SmolLM3 3B3B	Q8_0	3.73 GB	+44.3 GB	≤278Ceiling
Gemma 2 2B Instruct2B	Q8_0	3.34 GB	+44.7 GB	≤329Ceiling
Qwen3.5 2B2B	Q8_0	2.23 GB	+45.8 GB	≤436Ceiling
Qwen3 1.7B Instruct1.7B	Q8_0	2.39 GB	+45.6 GB	≤497Ceiling
Qwen2.5 1.5B Instruct1.5B	Q8_0	1.83 GB	+46.2 GB	≤557Ceiling
Llama 3.2 1B Instruct1B	Q8_0	1.51 GB	+46.5 GB	≤691Ceiling
Qwen3.5 0.8B0.8B	Q8_0	0.94 GB	+47.1 GB	≤1067Ceiling
Qwen2.5 0.5B Instruct0.5B	Q8_0	0.6 GB	+47.4 GB	≤1763Ceiling

How this list is built

Each row is the lowest-perplexity-loss quant of that model whose estimated total — weights plus KV cache at 4K context plus activation buffer — uses at most 88% of the card. That is the calculator's "green" threshold, so every row here has real headroom rather than only just fitting. Raise the context length and the list shortens; the calculator lets you check any combination directly. The speed column is this card: a measured run, a range scaled from this card's own Llama 2 7B benchmark, or a bandwidth ceiling. A figure that only exists for the RTX 4090 is labelled as that card.

A further 1 models load but with no headroom to spare (up to 105% of VRAM) — the Quant Hub’s GPU chips count those too, which is why its number is higher.

Open this card in the VRAM calculator Feel this speed Browse all 110 models

Measured on this card

No benchmark runs in this index were recorded on a Radeon PRO W7900 48G. Every figure on this page is calculated from the model architecture and the quant level — treat them as estimates, not measurements.

Setup guides for this card

Step-by-step guides written for the Radeon PRO W7900 48G's runtime or memory size, with the commands and the checks that prove the model is running on the card.

AMD GPU + llama.cpp via ROCm (Quick Start)

Common questions

What is the best local LLM for a Radeon PRO W7900 48G?

For everyday use, Nemotron 3.5 Bolt 74B-A9B at IQ2_XXS — about 23.1 GB of the card's 48 GB at 4K context, leaving 25.0 GB for a longer window. If you want the largest thing that will load, that is GLM-4.5-Air at Q2_K (37.2 GB, 10.8 GB spare). "Best" here means best fit for the memory budget — this index does not run task benchmarks, so it cannot tell you which model is smarter.

Can a Radeon PRO W7900 48G run Llama 4 Scout 17B (16E)?

No. Its smallest build here, Q3_K_M, needs about 55.9 GB and this card has 48.0 GB usable — short by 7.9 GB before any context beyond 4K. The largest model this card does clear is GLM-4.5-Air.

How many tokens per second does a Radeon PRO W7900 48G do on an 8B model at Q4?

This index has no measured run on this card, so it does not publish a figure. What can be stated from specifications: generating a token requires reading every weight once, Llama 3.1 8B Instruct at Q4_K_M is 4.6 GB of weights, and this card moves 864 GB/s — a ceiling near 187 tok/s. Batch size, context length, the runtime and how much of the model sits in cache all take you below it.

Radeon PRO W7900 48G or Radeon AI PRO R9700 32G for local LLMs?

They do not hold the same set: Radeon PRO W7900 48G fits 94 and Radeon AI PRO R9700 32G 88 of 110 at 4K. Speed is the other difference — 864 GB/s against 640 GB/s, a 1.4× gap in how fast the weights can be read, which is what token generation is bound by. The Radeon PRO W7900 48G generates faster on any model both can hold.

This page's figures change when the model or the runtime does.Last updated 2026-10-18 RSS → /feed.xml