Apple · apple

Apple M5

Apple M5 has 32 GB of unified memory at 154 GB/s — about 22.32 GiB usable after driver and compositor overhead. 1985 of 2118 indexed models fit at 8K context with q8_0 KV. Note only 24 GB of its 32 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
32 GB
LPDDR5X-9600
Bandwidth
154 GB/s
128-bit bus
Tensor FP16
—
dense
TDP
—
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1705vision language 176audio tts 21image 2video 16audio asr 39embedding 26

What fits at 8K context

largest quantization that fits, per model · 1985 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Laguna-S-2.1MoEIQ1_S118B23.15 GiB0.27 GiB24.00 GiB0.00 GiB25±37%
Huihui-GLM-4.7-Flash-abliterated-57BMoEI1-IQ3_M57.3B22.85 GiB0.56 GiB24.00 GiB0.00 GiB20±37%
command-r-35b-writer-v2IQ4_XS35.0B18.02 GiB5.31 GiB23.99 GiB0.01 GiB5±8.3%
Qwen2.5-Coder-14B-InstructQ6_K14.8B22.58 GiB0.80 GiB23.98 GiB0.02 GiB5±8.3%
InternVL3_5-30B-A3BQ6_K30.8B23.38 GiB0.00 GiB23.98 GiB0.02 GiB5±8.3%
Qwen3.5-35B-A3BMoEQ5_K_S36.0B23.33 GiB0.08 GiB23.97 GiB0.03 GiB29±37%
Qwen3.6-35B-A3BMoEQ5_K_S36.0B23.33 GiB0.08 GiB23.97 GiB0.03 GiB29±37%
Hunyuan-A13B-InstructMoEUD-IQ1_S80.4B22.89 GiB0.53 GiB23.97 GiB0.03 GiB5±8.3%
OmniAtlas-Qwen3-30B-A3BI1-Q6_K31.7B23.37 GiB0.00 GiB23.96 GiB0.04 GiB5±8.3%
Qwen3-Omni-30B-A3B-CaptionerI1-Q6_K31.7B23.37 GiB0.00 GiB23.96 GiB0.04 GiB5±8.3%
Hy-MT2-30B-A3BMoEQ6_K30.1B23.01 GiB0.40 GiB23.95 GiB0.05 GiB21±37%
Nemotron-Labs-Audex-30B-A3BQ4_K_L32.0B23.35 GiB0.00 GiB23.95 GiB0.05 GiB5±8.3%
Midnight-Miqu-70B-v1.5I1-Q2_K_S69.0B21.94 GiB1.33 GiB23.94 GiB0.06 GiB6±8.3%
Darwin-35B-A3B-OpusMoEQ5_K_M36.0B23.30 GiB0.08 GiB23.94 GiB0.06 GiB29±37%
Aurora-Code-1MoEQ5_K_M34.7B23.30 GiB0.08 GiB23.94 GiB0.06 GiB29±37%
grug-35b-v2MoEQ5_K_M35.1B23.30 GiB0.08 GiB23.94 GiB0.06 GiB29±37%
grug-35bMoEQ5_K_M35.1B23.30 GiB0.08 GiB23.94 GiB0.06 GiB29±37%
WorldSim-Opus-3.6-35B-A3BMoEQ5_K_M35.1B23.30 GiB0.08 GiB23.94 GiB0.06 GiB29±37%
Qwen3.6-35B-A3B-AnkoMoEQ5_K_M35.1B23.30 GiB0.08 GiB23.94 GiB0.06 GiB29±37%
KAT-Coder-V2.5-DevMoEQ5_K_M34.7B23.30 GiB0.08 GiB23.94 GiB0.06 GiB29±37%
Ornith-1.0-35BMoEQ5_K_M34.7B23.30 GiB0.08 GiB23.94 GiB0.06 GiB29±37%
Nex-N2-miniMoEQ5_K_M35.1B23.30 GiB0.08 GiB23.94 GiB0.06 GiB29±37%
EuroLLM-22B-Instruct-2512Q8_022.6B22.41 GiB0.90 GiB23.91 GiB0.09 GiB6±8.3%
Nemotron-Cascade-2-30B-A3BMoEQ4_K_L31.6B23.15 GiB0.22 GiB23.90 GiB0.10 GiB24±37%
Gemma-4-Novelist-Eclipse-31BQ5_K_M32.7B21.96 GiB1.29 GiB23.88 GiB0.12 GiB6±8.3%
Gemma-4-31B-StyleTuneQ5_K_M32.7B21.96 GiB1.29 GiB23.88 GiB0.12 GiB6±8.3%
Qwen-AgentWorld-35B-A3BMoEUD-Q5_K_S34.7B23.23 GiB0.08 GiB23.87 GiB0.13 GiB29±37%
solar-pro-preview-instructKV unresolvedQ8_022.1B21.91 GiB1.33 GiB23.85 GiB0.15 GiB6±8.3%
Qwen3-Coder-Next-REAMMoEI1-IQ3_XS60.3B23.19 GiB0.10 GiB23.83 GiB0.17 GiB30±37%
CallerQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
Dumpling-Qwen2.5-32BQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
OREAL-32BQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
openhands-lm-32b-v0.1Q5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
LongWriter-Zero-32BQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
OpenCodeReasoning-Nemotron-32B-IOIQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
Qwen2.5-Coder-32B-Instruct-abliteratedQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
OlympicCoder-32BQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
OpenCodeReasoning-Nemotron-32BQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
OpenThinker-32BQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
QwQ-32B-ArliAI-RpR-v4Q5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
Qwen2.5-Coder-32B-InstructQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
QwQ-32B-abliteratedQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
OpenThinker2-32BQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
INTELLECT-2Q5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
Qwen2.5-32B-InstructQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
QwQ-32B-PreviewQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
Qwen2.5-Coder-32BQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
Qwen2.5-32b-RP-InkQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
TinyR1-32B-PreviewQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
deepseek-r1-qwen-2.5-32B-ablatedQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
Rombos-LLM-V2.5-Qwen-32bQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
DeepSeek-R1-Distill-Qwen-32B-abliteratedQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
Qwen2.5-32B-ArliAI-RPMax-v1.3Q5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
DeepSeek-R1-Distill-Qwen-32BQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
Qwen2.5-VL-32B-InstructQ5_K_L33.5B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
EVA-Qwen2.5-32B-v0.2Q5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
EVA-Qwen2.5-32B-v0.1Q5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
cogito-v1-preview-qwen-32BQ5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
QwQ-32B-Snowdrop-v0Q5_K_L32.8B22.11 GiB1.06 GiB23.82 GiB0.18 GiB6±8.3%
Qwen3.5-122B-A10B-hereticMoEI1-IQ1_S123B23.13 GiB0.10 GiB23.81 GiB0.19 GiB29±37%
●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 Apple M5 run?
1985 of 2118 indexed open-weight models fit a Apple M5 at 8,192 context with q8_0 KV cache, the largest being Laguna-S-2.1 at IQ1_S. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M5 actually have?
Its nameplate is 32 GB, but about 22.32 GiB is available to a model once driver and compositor overhead is accounted for, and only 24 GB of the pool can be allocated to the GPU at all.
Is a Apple M5 fast for local AI?
Its memory bandwidth is 154 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.