What AI models can a Radeon Pro W7800 run?

2031 of 2118 indexed open-weight models fit a Radeon Pro W7800 at 32,768 context with q8_0 KV cache, the largest being Step-3.5-Flash-REAP-121B-A11B at I1-IQ2_M. That covers text, vision-language, image, video and speech models.

How much usable memory does a Radeon Pro W7800 actually have?

Its nameplate is 48 GB, but about 44.64 GiB is available to a model once driver and compositor overhead is accounted for.

Is a Radeon Pro W7800 fast for local AI?

Its memory bandwidth is 864 GB/s, and that figure — not teraflops — is what governs token generation speed. Capacity decides what you can run; bandwidth decides how fast it runs.

AMD · workstation

Radeon Pro W7800

Radeon Pro W7800 has 48 GB of VRAM at 864 GB/s — about 44.64 GiB usable after driver and compositor overhead. 2031 of 2118 indexed models fit at 32K context with q8_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed

Memory

48 GB

GDDR6

Bandwidth

864 GB/s

384-bit bus

Tensor FP16

—

dense

TDP

281 W

Context4K 8K 16K 32K 64K 128K

KV cachef16 q8_0 q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context

text 1743vision language 184image 2video 16audio tts 21embedding 26audio asr 39

What fits at 32K context

largest quantization that fits, per model · 2031 of 2118 indexed

Model	Best quant	Params	Weights●	KV●	Total in memory◐	Headroom◐	tok/s◐
Step-3.5-Flash-REAP-121B-A11B	I1-IQ2_M	121B	36.79 GiB	6.92 GiB	44.64 GiB	0.00 GiB	13±26.5%
GLM-4.7-Flash-REAP-23B-A3B-absolute-heresyMoE	BF16	23.0B	42.85 GiB	0.88 GiB	44.64 GiB	0.00 GiB	44±37%
GLM-4.7-Flash-REAP-23B-A3BMoE	BF16	23.0B	42.85 GiB	0.88 GiB	44.64 GiB	0.00 GiB	44±37%
GLM-4.5VMoE	I1-Q2_K	108B	40.61 GiB	3.05 GiB	44.59 GiB	0.05 GiB	38±37%
Assistant_Pepe_70B	Q4_K_S	70.6B	38.22 GiB	5.31 GiB	44.56 GiB	0.08 GiB	13±26.5%
Mistral-Medium-3.5-128B	IQ2_XXS	128B	37.65 GiB	5.84 GiB	44.56 GiB	0.08 GiB	13±26.5%
Qwen3.5-122B-A10BMoE	Q2_K	125B	43.21 GiB	0.40 GiB	44.53 GiB	0.11 GiB	68±37%
Phi-3.5-MoE-instructMoEKV unresolved	Q8_0	41.9B	41.44 GiB	2.13 GiB	44.48 GiB	0.16 GiB	32±37%
Qwen3.6-35B-A3B-REAM-160-ru-agentMoE	BF16	23.6B	43.09 GiB	0.33 GiB	44.33 GiB	0.31 GiB	59±37%
CalmeRys-78B-Orpo-v0.1	I1-Q3_K_M	78.0B	37.54 GiB	5.71 GiB	44.28 GiB	0.36 GiB	13±26.5%
calme-2.3-rys-78b	Q3_K_M	78.0B	37.54 GiB	5.71 GiB	44.28 GiB	0.36 GiB	13±26.5%
GLM-4.6VMoE	IQ2_M	108B	40.26 GiB	3.05 GiB	44.24 GiB	0.40 GiB	39±37%
ALIA-40b-fc-2606	Q8_0	40.4B	40.02 GiB	3.19 GiB	44.22 GiB	0.42 GiB	13±26.5%
ALIA-40b-instruct-2606	Q8_0	40.4B	40.02 GiB	3.19 GiB	44.22 GiB	0.42 GiB	13±26.5%
Qwen3-53B-A3B-2507-THINKING-TOTAL-RECALL-v2-MASTER-CODERMoE	I1-Q6_K	53.0B	40.54 GiB	2.79 GiB	44.22 GiB	0.42 GiB	37±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolved	Q2_K	109B	40.03 GiB	3.19 GiB	44.15 GiB	0.49 GiB	38±37%
NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16MoE	Q4_K_S	75.4B	43.15 GiB	0.00 GiB	44.12 GiB	0.52 GiB	102±37%
Wizard-Vicuna-30B-Uncensored	I1-Q4_K_S	32.5B	17.21 GiB	25.90 GiB	44.08 GiB	0.56 GiB	13±26.5%
archangel_sft-kto_llama30b	I1-Q4_K_S	32.5B	17.21 GiB	25.90 GiB	44.08 GiB	0.56 GiB	13±26.5%
GLM-4.5-Air-REAP-82B-A12BMoE	Q3_K_L	81.9B	40.10 GiB	3.05 GiB	44.08 GiB	0.56 GiB	35±37%
Huihui-Qwen3-Coder-Next-abliteratedMoE	Q4_0	79.7B	42.78 GiB	0.40 GiB	44.07 GiB	0.57 GiB	76±37%
Apertus-70B-Instruct-2509	Q4_K_S	70.6B	37.67 GiB	5.31 GiB	44.06 GiB	0.58 GiB	13±26.5%
WizardLM-Uncensored-SuperCOT-StoryTelling-30b	Q4_K_S	32.5B	17.17 GiB	25.90 GiB	44.04 GiB	0.60 GiB	13±26.5%
Mixtral_34Bx2_MoE_60BMoE	Q5_K_M	60.8B	39.03 GiB	3.98 GiB	44.00 GiB	0.64 GiB	7±37%
OYM-Qimi-122B-A10B-K2.6MoE	I1-Q2_K	125B	42.67 GiB	0.40 GiB	43.99 GiB	0.65 GiB	68±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoE	Q2_K	123B	42.66 GiB	0.40 GiB	43.99 GiB	0.65 GiB	68±37%
Meta-Llama-3-70B-Instruct	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.92 GiB	0.72 GiB	13±26.5%
Maenad-70B	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-Reasoner	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
calme-2.4-llama3-70b	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
calme-2.2-llama3-70b	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Rombos-LLM-70b-Llama-3.3	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
L3.3-Electra-R1-70b	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
L3.3-70B-Magnum-v4-SE	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Latxa-Llama-3.1-70B-Instruct-v2	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Llama-3.3_70_b_uncensored_continued	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Llama-3.3-70B-Instruct-abliterated	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Strawberrylemonade-L3-70B-v1.2	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
grok-oss-Revenant-70B	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Llama-3.1-Nemotron-70B-Instruct-HF	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
L3.3-70B-Euryale-v2.3	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Hermes-4-70B-heretic	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Hermes-4-70B	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Llama-3.3-70B-Instruct	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Llama-3.1-70B	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Hermes-3-Llama-3.1-70B	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Anubis-70B-v1.2	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Golem-70B-v1b	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
DeepSeek-R1-Distill-Llama-70B-abliterated	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
DeepSeek-R1-Distill-Llama-70B-heretic	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
DeepSeek-R1-Distill-Llama-70B	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
llama-3-firefunction-v2	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Legion-V2.1-LLaMa-70B	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Tess-R1-Limerick-Llama-3.1-70B	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
SEMIKONG-70B	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
functionary-medium-v3.2KV unresolved	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Infinity-Instruct-7M-Gen-Llama3_1-70B	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
New-Dawn-Llama-3-70B-32K-v1.0	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Meta-Llama-3-70B-Instruct-abliterated-v3.5	I1-Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%
Athene-70B	Q4_K_S	70.6B	37.58 GiB	5.31 GiB	43.91 GiB	0.73 GiB	13±26.5%

●From the file◐Predictedwhat these mean

Speed is modeled, not measured: decode is memory-bandwidth bound, so tokens per second is bytes read per token against achievable bandwidth. Mixture-of-experts models carry a wider band because only the routed experts are read each step, and few have been measured publicly.

Questions people ask

What AI models can a Radeon Pro W7800 run?: 2031 of 2118 indexed open-weight models fit a Radeon Pro W7800 at 32,768 context with q8_0 KV cache, the largest being Step-3.5-Flash-REAP-121B-A11B at I1-IQ2_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon Pro W7800 actually have?: Its nameplate is 48 GB, but about 44.64 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Radeon Pro W7800 fast for local AI?: Its memory bandwidth is 864 GB/s, and that figure — not teraflops — is what governs token generation speed. Capacity decides what you can run; bandwidth decides how fast it runs.