Mac Studio M3 Ultra 96GB — local LLM fit & speed

Apple · 96 GB LPDDR5 · 819 GB/s · 480 W · MSRP $3,999 · Search Amazon (affiliate)

The Mac Studio M3 Ultra 96GB is a Apple accelerator with 96 GB LPDDR5 and about 819 GB/s peak memory bandwidth. AI Hashrate estimates local LLM decode (batch 1) at 8K context: fit counts weights + KV cache + 1 GB runtime overhead — not weights alone. At Q4, about 47 curated models fit fully on this card; at FP16, about 37 fit. Top Q4 speeds: Qwen3-0.6B ≈ 651.5 tok/s (estimated, fits); gemma-3-1b-it ≈ 521.2 tok/s (estimated, fits); Llama-3.2-1B-Instruct ≈ 434.3 tok/s (estimated, fits); Qwen3-1.7B ≈ 260.6 tok/s (estimated, fits); Qwen3-30B-A3B ≈ 173.7 tok/s (estimated, fits). Relative ranking is more reliable than absolute tok/s. See methodology for the bandwidth formula and measured-anchor policy. Methodology.

Table: decode speed estimates (Q4 / FP16). Measured rows override estimates. MSRP is list price, not live retail. Context default 8K.

Model	Params (B)	Quant	tok/s	Fits?
Qwen3-0.6B	0.8	Q4	651.5 est.	Yes
gemma-3-1b-it	1.0	Q4	521.2 est.	Yes
Llama-3.2-1B-Instruct	1.2	Q4	434.3 est.	Yes
Qwen3-1.7B	2.0	Q4	260.6 est.	Yes
Qwen3-0.6B	0.8	FP16	179.2 est.	Yes
Qwen3-30B-A3B	30.5	Q4	173.7 est.	Yes
Qwen3-Coder-30B-A3B-Instruct	30.5	Q4	173.7 est.	Yes
Qwen3-30B-A3B-Instruct-2507	30.5	Q4	173.7 est.	Yes
Qwen3-Next-80B-A3B-Instruct	81.3	Q4	173.7 est.	Yes
Qwen3-Coder-Next	79.7	Q4	173.7 est.	Yes
Llama-3.2-3B-Instruct	3.2	Q4	162.9 est.	Yes
gpt-oss-20b	21.5	Q4	144.8 est.	Yes
gemma-3-1b-it	1.0	FP16	143.3 est.	Yes
Phi-4-mini-instruct	3.8	Q4	137.2 est.	Yes
Phi-4-mini-reasoning	3.8	Q4	137.2 est.	Yes
gemma-3-4b-it	4.3	Q4	133.6 est.	Yes
Qwen3-4B-Instruct-2507	4.0	Q4	130.3 est.	Yes
Qwen3-4B-Thinking-2507	4.0	Q4	130.3 est.	Yes
Llama-3.2-1B-Instruct	1.2	FP16	119.4 est.	Yes
MiniMax-M2.7	228.7	Q4	82.7 est.	No
Mistral-7B-Instruct-v0.3	7.2	Q4	72.4 est.	Yes
Qwen3-1.7B	2.0	FP16	71.7 est.	Yes
Qwen2.5-7B-Instruct	7.6	Q4	68.6 est.	Yes
Qwen2.5-Coder-7B-Instruct	7.6	Q4	68.6 est.	Yes
DeepSeek-R1-Distill-Qwen-7B	7.6	Q4	68.6 est.	Yes
Llama-3.1-8B-Instruct	8.0	Q4	65.1 est.	Yes
gpt-oss-120b	120.4	Q4	65.1 est.	Yes
Qwen3-8B	8.2	Q4	63.6 est.	Yes
DeepSeek-R1-0528-Qwen3-8B	8.2	Q4	63.6 est.	Yes
Qwen3-Coder-Next	79.7	FP16	48.5 est.	No
Qwen3-30B-A3B	30.5	FP16	47.8 est.	Yes
Qwen3-Coder-30B-A3B-Instruct	30.5	FP16	47.8 est.	Yes
Qwen3-30B-A3B-Instruct-2507	30.5	FP16	47.8 est.	Yes
Qwen3-Next-80B-A3B-Instruct	81.3	FP16	46.7 est.	No
Llama-3.2-3B-Instruct	3.2	FP16	44.8 est.	Yes
gemma-3-12b-it	12.2	Q4	43.4 est.	Yes
GLM-4.7-Flash	31.2	Q4	43.4 est.	Yes
GLM-4.5-Air	110.5	Q4	43.4 est.	Yes
Mixtral-8x7B-Instruct-v0.1	46.7	Q4	40.4 est.	Yes
DeepSeek-V4-Flash	158.1	Q4	40.1 est.	Yes
gpt-oss-20b	21.5	FP16	39.8 est.	Yes
Phi-4-mini-instruct	3.8	FP16	37.7 est.	Yes
Phi-4-mini-reasoning	3.8	FP16	37.7 est.	Yes
gemma-3-4b-it	4.3	FP16	36.8 est.	Yes
Qwen3-4B-Instruct-2507	4.0	FP16	35.8 est.	Yes
Qwen3-4B-Thinking-2507	4.0	FP16	35.8 est.	Yes
Phi-4	14.7	Q4	35.5 est.	Yes
Phi-4-reasoning	14.7	Q4	35.5 est.	Yes
Qwen3-14B	14.8	Q4	35.2 est.	Yes
Qwen2.5-14B-Instruct	14.8	Q4	35.2 est.	Yes
Qwen2.5-Coder-14B-Instruct	14.8	Q4	35.2 est.	Yes
DeepSeek-R1-Distill-Qwen-14B	14.8	Q4	35.2 est.	Yes
Qwen3-235B-A22B	235.1	Q4	34.4 est.	No
Qwen3-235B-A22B-Instruct-2507	235.1	Q4	34.4 est.	No
Mixtral-8x7B-Instruct-v0.1	46.7	FP16	31.1 est.	No
Devstral-Small-2507	23.6	Q4	22.1 est.	Yes
Mistral-Small-3.2-24B-Instruct-2506	24.0	Q4	21.7 est.	Yes
Magistral-Small-2509	24.0	Q4	21.7 est.	Yes
Mistral-7B-Instruct-v0.3	7.2	FP16	19.9 est.	Yes
gemma-3-27b-it	27.4	Q4	19.2 est.	Yes
Qwen2.5-7B-Instruct	7.6	FP16	18.9 est.	Yes
Qwen2.5-Coder-7B-Instruct	7.6	FP16	18.9 est.	Yes
DeepSeek-R1-Distill-Qwen-7B	7.6	FP16	18.9 est.	Yes
Llama-3.1-8B-Instruct	8.0	FP16	17.9 est.	Yes
Qwen3-8B	8.2	FP16	17.5 est.	Yes
DeepSeek-R1-0528-Qwen3-8B	8.2	FP16	17.5 est.	Yes
Qwen3-32B	32.8	Q4	15.9 est.	Yes
Qwen2.5-32B-Instruct	32.8	Q4	15.9 est.	Yes
Qwen2.5-Coder-32B-Instruct	32.8	Q4	15.9 est.	Yes
DeepSeek-R1-Distill-Qwen-32B	32.8	Q4	15.9 est.	Yes
gemma-3-12b-it	12.2	FP16	11.9 est.	Yes
GLM-4.7-Flash	31.2	FP16	11.9 est.	Yes
GLM-4.5	358.3	Q4	10.3 est.	No
Phi-4	14.7	FP16	9.8 est.	Yes
Phi-4-reasoning	14.7	FP16	9.8 est.	Yes
Qwen3-14B	14.8	FP16	9.7 est.	Yes
Qwen2.5-14B-Instruct	14.8	FP16	9.7 est.	Yes
Qwen2.5-Coder-14B-Instruct	14.8	FP16	9.7 est.	Yes
DeepSeek-R1-Distill-Qwen-14B	14.8	FP16	9.7 est.	Yes
gpt-oss-120b	120.4	FP16	8.0 est.	No
Llama-3.1-70B-Instruct	70.6	Q4	7.4 est.	Yes
Llama-3.3-70B-Instruct	70.6	Q4	7.4 est.	Yes
DeepSeek-R1-Distill-Llama-70B	70.6	Q4	7.4 est.	Yes
Qwen2.5-72B-Instruct	72.7	Q4	7.2 est.	Yes
GLM-4.5-Air	110.5	FP16	6.2 est.	No
Devstral-Small-2507	23.6	FP16	6.1 est.	Yes
Mistral-Small-3.2-24B-Instruct-2506	24.0	FP16	6.0 est.	Yes
Magistral-Small-2509	24.0	FP16	6.0 est.	Yes
gemma-3-27b-it	27.4	FP16	5.3 est.	Yes
Qwen3-Coder-480B-A35B-Instruct	480.2	Q4	5.3 est.	No
Qwen3-32B	32.8	FP16	4.4 est.	Yes
Qwen2.5-32B-Instruct	32.8	FP16	4.4 est.	Yes
Qwen2.5-Coder-32B-Instruct	32.8	FP16	4.4 est.	Yes
DeepSeek-R1-Distill-Qwen-32B	32.8	FP16	4.4 est.	Yes
MiniMax-M1	456.1	Q4	4.4 est.	No
DeepSeek-V4-Flash	158.1	FP16	2.8 est.	No
Llama-3.1-70B-Instruct	70.6	FP16	2.2 est.	No
Llama-3.3-70B-Instruct	70.6	FP16	2.2 est.	No
DeepSeek-R1-Distill-Llama-70B	70.6	FP16	2.2 est.	No
Qwen2.5-72B-Instruct	72.7	FP16	2.0 est.	No
Llama-3.1-405B-Instruct	405.9	Q4	0.4 est.	No