M1 Ultra vs Radeon RX 6950 XT for local AI

Specs that matter

	M1 Ultra	Radeon RX 6950 XT
Memory compared	64 GB (base configuration)	16 GB
Memory bandwidth	800 GB/s	576 GB/s
Memory type	Unified (shared with the system)	Dedicated VRAM
Backend	Metal	Vulkan
Released	2022	2022

Every model on both

Grade and output speed at Q4_K_M with an 8K context; green marks the better of the two.

Model	ECI	M1 Ultra	Radeon RX 6950 XT
Kimi K3	157	F—	F—
DeepSeek V4 Pro 0813	155	F—	F—
DeepSeek V4.1 Flash	155	F—	F—
DeepSeek V4 Flash 0731	154	F—	F—
GLM 5.3 Flash	152	F—	F—
GLM 5.2	152	F—	F—
Kimi K2.6	151	F—	F—
tInkling Small	150	F—	F—
Kimi K2.7 Code	150	F—	F—
GLM 5.1	150	F—	F—
Qwen3.8 27B	149	A~34 t/s	C~15 t/s
DeepSeek V4 Pro	149	F—	F—
tInkling	149	F—	F—
Kimi K2.5	148	F—	F—
MiniMax M3	147	F—	F—
MiniMax M2.5	147	F—	F—
Qwen3.5 397B A17B	147	F—	F—
Qwen3.6 27B	147	A~35 t/s	C~17 t/s
DeepSeek V3.2	146	F—	F—
Nemotron 3 Ultra 550B A55B	146	F—	F—
DeepSeek V4 Flash	146	F—	F—
Kimi K2 Thinking	146	F—	F—
MiniMax M2.7	146	F—	F—
GLM 5	146	F—	F—
DeepSeek V3.2 Exp	145	F—	F—
Qwen3.6 35B A3B	144	S~126 t/s	C~65 t/s
Qwen3 235B A22B Thinking 2507	144	F—	F—
GLM 4.7	144	F—	F—
Gemma 4 31B	143	A~32 t/s	C~10 t/s
Qwen3.5 35B A3B	143	S~126 t/s	C~66 t/s
Gemma 4 26B A4B	142	S~110 t/s	C~76 t/s
Mistral Medium 3.5 128B	141	F—	F—
Kimi K2	140	F—	F—
gpt-oss 120b	140	F—	F—
DeepSeek V3.1	140	F—	F—
Qwen3 30B A3B Thinking 2507	140	S~116 t/s	C~68 t/s
Qwen3.5 9B	139	S~76 t/s	S~63 t/s
Qwen3 235B A22B 2507	139	F—	F—
gpt-oss 20b	138	S~122 t/s	S~121 t/s
Qwen3 30B A3B 2507	137	S~116 t/s	C~68 t/s
Magistral Small 2506	133	A~39 t/s	B~29 t/s
Mistral Small 3.2 24B	132	A~39 t/s	B~29 t/s
Magistral Small 2509	131	A~39 t/s	B~29 t/s
Phi-4 14B	130	S~54 t/s	S~42 t/s
Llama 4 Scout 17B 16E	130	F—	F—
Llama 3.3 70B	127	B~16 t/s	F—
Llama 3.1 8B	117	S~79 t/s	S~66 t/s
Command A+	—	F—	F—
Qwen3.8 Flash Next	—	F—	F—
Qwen3.5 122B A10B	—	F—	F—
Mistral Small 4 119B	—	F—	F—
Qwen3 Coder Next	—	B~126 t/s	F—
Kimi Linear 48B A3B	—	S~127 t/s	C~51 t/s
Olmo 3.1 32B Think	—	A~30 t/s	C~9 t/s
Nemotron 3.5 Lightning 30B A3B	—	S~116 t/s	C~47 t/s
GLM 4.7 Flash	—	S~125 t/s	C~83 t/s
North Mini Code 1.0	—	S~119 t/s	C~72 t/s
Granite 4.2 30B	—	A~32 t/s	C~10 t/s
Devstral Small 2 24B	—	A~39 t/s	B~29 t/s
Ministral 3 14B	—	S~58 t/s	S~45 t/s
Gemma 4 12B	—	S~65 t/s	S~52 t/s
Ministral 3 8B	—	S~76 t/s	S~64 t/s
Granite 4.2 8B	—	S~73 t/s	S~60 t/s
Gemma 4 E4B	—	S~82 t/s	S~70 t/s
Olmo 3 7B	—	S~70 t/s	S~57 t/s
Gemma 4 E2B	—	S~104 t/s	S~96 t/s
Qwen3.5 4B	—	S~109 t/s	S~102 t/s
Ministral 3 3B	—	S~113 t/s	S~108 t/s
Granite 4.2 3B	—	S~112 t/s	S~106 t/s
Llama 3.2 3B	—	S~115 t/s	S~110 t/s
SmolLM3 3B	—	S~120 t/s	S~118 t/s
LFM2.5 2.6B	—	S~130 t/s	S~132 t/s
Qwen3.5 2B	—	S~140 t/s	S~149 t/s
Qwen3.5 0.8B	—	S~163 t/s	S~193 t/s

Estimates from published specs and measurements, calibrated per backend. Your drivers, settings and runtime change the real numbers.