Intel · workstation

Arc Pro B60 24GB

Arc Pro B60 24GB has 24 GB of VRAM at 456 GB/s — about 22.32 GiB usable after driver and compositor overhead. 1872 of 2118 indexed models fit at 32K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
24 GB
GDDR6
Bandwidth
456 GB/s
192-bit bus
Tensor FP16
99 TF
dense
TDP
200 W
$599 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1597vision language 171audio asr 39audio tts 21image 2video 16embedding 26

What fits at 32K context

largest quantization that fits, per model · 1872 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Yi-34B-200K-DARE-megamerge-v8I1-Q3_K_S34.4B13.93 GiB7.50 GiB22.32 GiB0.00 GiB12±30%
dolphin-2.9.1-yi-1.5-34b-hereticQ3_K_S34.4B13.93 GiB7.50 GiB22.32 GiB0.00 GiB12±30%
dolphin-2.9.1-yi-1.5-34bI1-Q3_K_S34.4B13.93 GiB7.50 GiB22.32 GiB0.00 GiB12±30%
OrionStar-Yi-34B-Chat-LlamaI1-Q3_K_S34.4B13.93 GiB7.50 GiB22.32 GiB0.00 GiB12±30%
Yi-34B-200K-LlamafiedI1-Q3_K_S34.4B13.93 GiB7.50 GiB22.32 GiB0.00 GiB12±30%
Yi-1.5-34BQ3_K_S34.4B13.93 GiB7.50 GiB22.32 GiB0.00 GiB12±30%
Nous-Hermes-2-Yi-34BQ3_K_S34.4B13.93 GiB7.50 GiB22.32 GiB0.00 GiB12±30%
Merged-RP-Stew-V2-34BI1-Q3_K_S34.4B13.93 GiB7.50 GiB22.32 GiB0.00 GiB12±30%
Capybara-Tess-Yi-34B-200KQ3_K_S34.4B13.93 GiB7.50 GiB22.32 GiB0.00 GiB12±30%
Nous-Capybara-limarpv3-34BQ3_K_S34.4B13.93 GiB7.50 GiB22.32 GiB0.00 GiB12±30%
TildeOpen-30B-Instruct-LVI1-Q3_K_M30.7B13.93 GiB7.50 GiB22.31 GiB0.01 GiB12±30%
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16UD-Q4_K_S33.0B21.47 GiB0.00 GiB22.31 GiB0.01 GiB12±30%
CallerQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
Dumpling-Qwen2.5-32BQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
OREAL-32BQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
Baichuan-M2-32B-abliteratedQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
QwQ-32B-Preview-abliterated-linear25I1-Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
openhands-lm-32b-v0.1I1-Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
Qwen2.5-Coder-32B-abliteratedI1-Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
INTELLECT-2Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
LongWriter-Zero-32BQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
m1-32bI1-Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
XMainframe-v2-Instruct-32bI1-Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
Qwen2.5-Coder-32B-Python-SpecialistI1-Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
Qwen2.5-32b-RP-InkI1-Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
OpenCodeReasoning-Nemotron-32B-IOIQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
Qwen2.5-Coder-32B-Instruct-abliteratedQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
OlympicCoder-32BQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
OpenCodeReasoning-Nemotron-32BQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
OpenThinker-32BQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
QwQ-32B-ArliAI-RpR-v4Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
Qwen2.5-Coder-32B-InstructQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
Qwen2.5-Coder-32BQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
QwQ-32B-abliteratedQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
DeepSeek-R1-Distill-Qwen-32B-hereticI1-Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
InnoSpark-HPC-RM-32BI1-Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
OpenThinker2-32BQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
Qwen2.5-32B-InstructQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
Qwen2.5-Coder-32B-Instruct-UncensoredI1-Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
QwQ-32B-PreviewQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
DeepSeek-R1-Distill-Qwen-32B-abliteratedQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
TinyR1-32B-PreviewQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
deepseek-r1-qwen-2.5-32B-ablatedQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
Rombos-LLM-V2.5-Qwen-32bQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
QwQ-32BQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
Qwen2.5-32B-ArliAI-RPMax-v1.3Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
DeepSeek-R1-Distill-Qwen-32B-Blunt-UncensoredQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
DeepSeek-R1-Distill-Qwen-32BQ3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
Qwen2.5-VL-32B-InstructQ3_K_S33.5B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
EVA-Qwen2.5-32B-v0.2Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
EVA-Qwen2.5-32B-v0.1Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
cogito-v1-preview-qwen-32BI1-Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
QwQ-32B-Snowdrop-v0I1-Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
DeepSeek-R1-Distill-Qwen-32B-UncensoredI1-Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
RoguePlanet-DeepSeek-R1-Qwen-32B-RPI1-Q3_K_S32.8B13.40 GiB8.00 GiB22.30 GiB0.02 GiB12±30%
OpenBuddy-R1-0528-Distill-Qwen3-32B-Preview0-QATQ3_K_S32.8B13.40 GiB8.00 GiB22.29 GiB0.03 GiB12±30%
Qwen3-VL-32B-Instruct-ultra-uncensored-hereticI1-Q3_K_S33.4B13.40 GiB8.00 GiB22.29 GiB0.03 GiB12±30%
Huihui-Qwen3-VL-32B-Instruct-abliteratedI1-Q3_K_S33.4B13.40 GiB8.00 GiB22.29 GiB0.03 GiB12±30%
KAT-DevQ3_K_S32.8B13.40 GiB8.00 GiB22.29 GiB0.03 GiB12±30%
ColorGUI-32BI1-Q3_K_S33.4B13.40 GiB8.00 GiB22.29 GiB0.03 GiB12±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 B60 24GB run?
1872 of 2118 indexed open-weight models fit a Arc Pro B60 24GB at 32,768 context with f16 KV cache, the largest being Yi-34B-200K-DARE-megamerge-v8 at I1-Q3_K_S. That covers text, vision-language, image, video and speech models.
How much usable memory does a Arc Pro B60 24GB actually have?
Its nameplate is 24 GB, but about 22.32 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Arc Pro B60 24GB fast for local AI?
Its memory bandwidth is 456 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.