Hardware

Apple M3 Ultra

Apple M3 Ultra has 96 / 256 / 512 GB Unified and 819 GB/s. See which open models fit, at which quant, and roughly how fast. Figures below use the default memory (96 GB). Updated 2026-10-17.

Memory96 / 256 / 512 GB Unified

Bandwidth819 GB/s

Kindmac

Year2025

Apple quotes 819 GB/s and up to 512GB unified memory. Source.

RSS for new models on this device

Models that fit

Model	Verdict	Quant	Context	tok/s
gpt oss 120b 117B	Tight fit	UD-Q8_K_XL	8K	37–50
Mistral Small 4 119B 2603 119B	Tight fit	UD-Q4_K_XL	8K	33–45
GLM 4.5 Air 110B	Tight fit	UD-Q4_K_XL	8K	28–39
Llama 4 Scout 17B 16E Instruct 108B	Runs	UD-Q4_K_XL	8K	30–42
GLM 4.6V 107B	Tight fit	UD-Q4_K_XL	8K	29–40
Qwen3.5 122B A10B MTP 125B	Tight fit	UD-Q4_K_S	8K	74–101
NVIDIA Nemotron 3 Super 120B A12B 121B	Tight fit	IQ4_NL	8K	63–86
Qwen3 Coder Next 79.7B	Tight fit	UD-Q6_K_XL	8K	34–47
Qwen3 Next 80B A3B Instruct 79.7B	Tight fit	UD-Q6_K_XL	8K	133–181
Qwen3 Next 80B A3B Thinking 79.7B	Tight fit	UD-Q6_K_XL	8K	133–181
Step 3.7 Flash 197B	Tight fit	UD-Q2_K_XL	4K	25–35
MiniMax M2 229B	Tight fit	UD-IQ1_M	8K	31–42
Qwen3.8 Flash Next 177B	Tight fit	UD-IQ1_S	8K	36–49
Qwen2.5 Coder 32B Instruct 32.8B	Runs great	Q8_0	8K	12–17
Kwaipilot KAT Coder V2.5 Dev 34.7B	Tight fit	BF16	8K	35–48
Qwen3.8 35B A3B Distill 35.5B	Tight fit	BF16	8K	73–100
Qwen3.6 35B A3B 34.7B	Tight fit	BF16	8K	73–100
Qwen3.5 35B A3B 34.7B	Tight fit	BF16	8K	73–100
NVIDIA Nemotron 3.5 Lightning 30B A3B 32.9B	Tight fit	BF16	8K	70–95
NVIDIA Nemotron 3 Nano Omni 30B A3B Reasoning 31.6B	Runs great	UD-Q8_K_XL	8K	119–162
Nemotron 3 Nano 30B A3B 31.6B	Runs	BF16	8K	70–95
GLM 4.7 Flash 29.9B	Runs	BF16	8K	34–47
Qwen3 Coder 30B A3B Instruct 30.5B	Runs	BF16	8K	66–90
Qwen3 30B A3B Instruct 2507 30.5B	Runs	BF16	8K	66–90
Qwen3 VL 30B A3B Instruct 30.5B	Runs	BF16	8K	66–90
Qwen3 30B A3B Thinking 2507 30.5B	Runs	BF16	8K	66–90
gemma 4 26B A4B it 25.2B	Runs	BF16	8K	45–61
LFM2.5 24B A2B 24.0B	Runs	BF16	8K	96–131
GLM 4.7 Flash REAP 23B A3B 23.0B	Runs	BF16	8K	70–95
gpt oss 20b 20.9B	Runs great	UD-Q8_K_XL	8K	112–152
Qwen2.5 14B Instruct 14.8B	Runs great	F16	8K	14–20
Qwen3 14B 14.8B	Runs great	BF16	8K	15–20
NousResearch Hermes 4 14B 14.8B	Runs great	BF16	8K	15–20
DeepSeek R1 Distill Qwen 14B 14.8B	Runs great	F16	8K	14–20
Qwen2.5 Coder 14B Instruct 128K 14.8B	Runs great	F16	8K	14–20
phi 4 14.7B	Runs great	F16	8K	15–20
Ministral 3 14B Instruct 2512 13.5B	Runs great	BF16	8K	16–22
Ministral 3 14B Reasoning 2512 13.5B	Runs great	BF16	8K	16–22
gemma 4 12b it 11.9B	Runs great	BF16	8K	17–23
GLM 4.6V Flash 9.4B	Runs great	BF16	8K	24–32

90 more models in the catalog do not fit this default profile, or ranked below the first 40 that do.

Example: Tight fit fully in unified memory on your Apple M3 Ultra at UD-Q8_K_XL, 8K context, about 37–50 tok/s. Weights 60.0 GB plus 0.29 GB of KV cache — 61.2 GB of 72.0 GB usable. Up to 128K context still fits. Weights shown in the verdict use GiB (the same 1024³ counting as GPU memory).