Intel · workstation

Arc Pro B50 16GB

Arc Pro B50 16GB has 16 GB of VRAM at 224 GB/s — about 14.88 GiB usable after driver and compositor overhead. 1858 of 2118 indexed models fit at 16K context with q4_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
16 GB
GDDR6
Bandwidth
224 GB/s
128-bit bus
Tensor FP16
85 TF
dense
TDP
70 W
$349 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1593video 15vision language 162embedding 26audio asr 39audio tts 21image 2

What fits at 16K context

largest quantization that fits, per model · 1858 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Laguna-XS-2.1MoEQ3_K_S33.4B13.87 GiB0.21 GiB14.88 GiB0.00 GiB45±37%
Qwen3-Coder-REAP-25B-A3BMoEQ4_K_S24.9B13.66 GiB0.42 GiB14.88 GiB0.00 GiB30±37%
Teuken-7B-instruct-research-v0.4F167.5B13.89 GiB0.14 GiB14.87 GiB0.01 GiB9±30%
Qwen3.6-27B-A3B-CoderMoEI1-Q4_026.7B13.97 GiB0.09 GiB14.87 GiB0.01 GiB40±37%
Skyfall-31B-v4.2Q3_K_S31.4B13.00 GiB0.95 GiB14.87 GiB0.01 GiB9±30%
Rocinante-XL-16B-v1Q6_K_L16.1B13.06 GiB0.95 GiB14.85 GiB0.03 GiB9±30%
Qwen-AgentWorld-35B-A3BMoEUD-IQ3_S34.7B13.96 GiB0.09 GiB14.85 GiB0.03 GiB48±37%
Ornith-1.0-35BMoEUD-IQ3_S34.7B13.96 GiB0.09 GiB14.85 GiB0.03 GiB48±37%
granite-4.0-h-smallMoEIQ3_M32.2B13.99 GiB0.07 GiB14.85 GiB0.03 GiB25±37%
OLMo-2-1124-13B-InstructQ6_K13.7B10.48 GiB3.52 GiB14.85 GiB0.03 GiB9±30%
LFM2-24B-A2BMoEQ4_123.8B13.93 GiB0.09 GiB14.83 GiB0.05 GiB38±37%
Qwen3.5-27B-Engineer-Deckard-GeminiI1-IQ4_XS27.7B13.68 GiB0.28 GiB14.82 GiB0.06 GiB9±30%
Qwen3.5-27B-HERETIC-Polaris-Advanced-Thinking-Alpha-uncensoredI1-IQ4_XS27.4B13.68 GiB0.28 GiB14.82 GiB0.06 GiB9±30%
Qwen3.5-27B-Deckard-PKD-Heretic-Uncensored-ThinkingI1-IQ4_XS27.4B13.68 GiB0.28 GiB14.82 GiB0.06 GiB9±30%
Huihui-Qwen3.5-27B-abliteratedI1-IQ4_XS27.8B13.68 GiB0.28 GiB14.82 GiB0.06 GiB9±30%
Qwen3.5-27B-Unredacted-MAXI1-IQ4_XS27.4B13.68 GiB0.28 GiB14.82 GiB0.06 GiB9±30%
Qwen3.5-27B-hereticI1-IQ4_XS27.4B13.68 GiB0.28 GiB14.82 GiB0.06 GiB9±30%
Qwen3.5-27B-DerestrictedI1-IQ4_XS27.8B13.68 GiB0.28 GiB14.82 GiB0.06 GiB9±30%
Qwen3.5-27B-Claude-4.6-Opus-Reasoning-DistilledI1-IQ4_XS27.8B13.68 GiB0.28 GiB14.82 GiB0.06 GiB9±30%
North-Mini-Code-1.0MoEIQ3_M30.5B13.84 GiB0.20 GiB14.82 GiB0.06 GiB36±37%
Noromaid-20b-v0.1.1I1-IQ3_M20.0B8.53 GiB5.45 GiB14.82 GiB0.06 GiB9±30%
Trinity-MiniMoEIQ4_NL26.1B13.92 GiB0.10 GiB14.82 GiB0.06 GiB38±37%
internlm2-math-plus-20bI1-Q5_K_M19.9B13.11 GiB0.84 GiB14.81 GiB0.07 GiB9±30%
gemma-4-E4B-uncensoredF167.9B13.92 GiB0.08 GiB14.81 GiB0.07 GiB9±30%
gemma-4-E4B-it-qat-heretic_decensoredF167.9B13.92 GiB0.08 GiB14.81 GiB0.07 GiB9±30%
gemma-4-E4B-it-QAT-SOMPOA-heresyF167.9B13.92 GiB0.08 GiB14.81 GiB0.07 GiB9±30%
gemma-4-E4B-it-hereticBF168.0B13.92 GiB0.08 GiB14.81 GiB0.07 GiB9±30%
Tinman-gemma4-companion-mergedBF167.9B13.92 GiB0.08 GiB14.81 GiB0.07 GiB9±30%
EXAONE-4.0-32BIQ3_M32.0B13.39 GiB0.52 GiB14.81 GiB0.07 GiB9±30%
Wan2.2-S2V-14BQ5_K_M16.3B13.97 GiB0.00 GiB14.81 GiB0.07 GiB9±30%
IQuest-Coder-V1-40B-InstructI1-IQ2_M39.8B12.50 GiB1.41 GiB14.81 GiB0.07 GiB9±30%
Gemma4-Gutenberg-31BIQ3_XS31.3B12.89 GiB1.03 GiB14.80 GiB0.08 GiB9±30%
gemma-4-31B-itIQ3_XS31.3B12.89 GiB1.03 GiB14.80 GiB0.08 GiB9±30%
Gemma4-Gutenberg-31B-HereticIQ3_XS31.3B12.89 GiB1.03 GiB14.80 GiB0.08 GiB9±30%
Equinox-31BIQ3_XS31.3B12.89 GiB1.03 GiB14.80 GiB0.08 GiB9±30%
gemma-4-31B-it-SDFT-Heretic-RPIQ3_XS30.7B12.89 GiB1.03 GiB14.80 GiB0.08 GiB9±30%
MN-GRAND-23.5B-Gutenberg-UNCENSORED-V2-GLM4.7-ThinkingI1-Q4_K_S23.4B12.53 GiB1.42 GiB14.80 GiB0.08 GiB9±30%
GPT-NeoX-20B-ErebusI1-IQ3_M20.6B9.27 GiB4.64 GiB14.80 GiB0.08 GiB9±30%
Qwen3-Coder-30B-A3B-InstructMoEQ3_K_L30.5B13.58 GiB0.42 GiB14.80 GiB0.08 GiB33±37%
Qwen3-VL-30B-A3B-ThinkingMoEQ3_K_L31.1B13.58 GiB0.42 GiB14.80 GiB0.08 GiB33±37%
MiroThinker-v1.0-30BMoEQ3_K_L30.5B13.58 GiB0.42 GiB14.80 GiB0.08 GiB33±37%
Qwen3-30B-A3BMoEQ3_K_L30.5B13.58 GiB0.42 GiB14.80 GiB0.08 GiB33±37%
Pantheon-Proto-RP-1.8-30B-A3BMoEQ3_K_L30.5B13.58 GiB0.42 GiB14.80 GiB0.08 GiB33±37%
Tongyi-DeepResearch-30B-A3BMoEQ3_K_L30.5B13.58 GiB0.42 GiB14.79 GiB0.09 GiB33±37%
CallerIQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
Dumpling-Qwen2.5-32BIQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
OREAL-32BIQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
QwQ-32B-Preview-abliterated-linear25I1-IQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
openhands-lm-32b-v0.1I1-IQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
Qwen2.5-Coder-32B-abliteratedI1-IQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
m1-32bI1-IQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
XMainframe-v2-Instruct-32bI1-IQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
Qwen2.5-Coder-32B-Python-SpecialistI1-IQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
Qwen2.5-32b-RP-InkI1-IQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
LongWriter-Zero-32BIQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
OpenCodeReasoning-Nemotron-32B-IOIIQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
Qwen2.5-Coder-32B-Instruct-abliteratedIQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
OlympicCoder-32BIQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
OpenCodeReasoning-Nemotron-32BIQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
OpenThinker-32BIQ3_XS32.8B12.76 GiB1.13 GiB14.79 GiB0.09 GiB9±30%
●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 Arc Pro B50 16GB run?
1858 of 2118 indexed open-weight models fit a Arc Pro B50 16GB at 16,384 context with q4_0 KV cache, the largest being Laguna-XS-2.1 at Q3_K_S. That covers text, vision-language, image, video and speech models.
How much usable memory does a Arc Pro B50 16GB actually have?
Its nameplate is 16 GB, but about 14.88 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Arc Pro B50 16GB fast for local AI?
Its memory bandwidth is 224 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.