Which local LLMs can the Apple M4 Max (32-core GPU; MacBook Pro, Mac Studio) run?

Unified memory — by default the GPU can use about 70% of it

Assumes 8k context and an f16 KV cache. Verdicts are shown for each memory size.

Specs

Memory options: 36 GB
Memory bandwidth: 410.0 GB/s
FP16 compute: 12.9 TFLOPS
Launch year: 2024
Macs with this chip: MacBook Pro: 36 GB
Mac Studio: 36 GB
Also written as: M4 Max 32

Verdicts at a glance

At Q4_K_M (or the closest available quant) with 8k context, with 36 GB of memory.

Runs great34 modelsRuns well14 modelsRuns slowly1 modelWon't run10 models2 more models run at a lower quant.

How to read the verdicts

Runs great: Fully on the GPU at 20 tok/s or more
Runs well: Fully on the GPU at 8–20 tok/s; MoE experts in system RAM at 20 tok/s or more; or at least 90% on the GPU at 8 tok/s or more
Runs slowly: 2–8 tok/s; CPU-only; less than 90% on the GPU; or MoE experts in system RAM below 20 tok/s
Won't run: Does not fit, or under 2 tok/s

Every model on the Apple M4 Max (32-core GPU)

Scroll sideways to see every column.

59 models sorted by verdict and speed, with 36 GB of memory
Model	Verdict	Speed	Quant	Memory	Runs as	Context	Notes
MiniCPM5 1B	Runs great	est. 349.6 tok/s279.7–419.5calibrated estimate ±20%	Q4_K_M	1.5 / 25.2 GB	Unified memory	up to 128k	thinking mode
Llama 3.2 1B	Runs great	est. 288.9 tok/s231.1–346.7calibrated estimate ±20%	Q4_K_M	1.7 / 25.2 GB	Unified memory	up to 128k	—
EXAONE 4.0 1.2B	Runs great	est. 237.3 tok/s189.8–284.7calibrated estimate ±20%	Q4_K_M	1.9 / 25.2 GB	Unified memory	up to 64k	thinking mode
Ling-3.0 Tiny	Runs great	est. 227.4 tok/s159.2–295.6theoretical estimate ±30%	Q4_K_M	5.5 / 25.2 GB	Unified memory	up to 128k	thinking mode
LFM2.5 2.6B	Runs great	est. 172.7 tok/s120.9–224.5theoretical estimate ±30%	Q4_K_M	2.4 / 25.2 GB	Unified memory	up to 128k	thinking mode
HyperCLOVA X SEED 1.5B	Runs great	est. 171.7 tok/s137.3–206.0calibrated estimate ±20%	Q4_K_M	2.4 / 25.2 GB	Unified memory	up to 16k	—
MiniCPM5 2B	Runs great	est. 162.9 tok/s130.4–195.5calibrated estimate ±20%	Q4_K_M	2.5 / 25.2 GB	Unified memory	up to 128k	thinking mode
Gemma 4 E2B	Runs great	est. 148.7 tok/s104.1–193.3theoretical estimate ±30%	Q4_K_M	3.7 / 25.2 GB	Unified memory	up to 128k	thinking mode
Mellum2.1 12B-A2.5B Thinking	Runs great	est. 137.5 tok/s110.0–165.0calibrated estimate ±20%	Q4_K_M	8.8 / 25.2 GB	Unified memory	up to 128k	thinking mode
Spark-X2.5 4B	Runs great	est. 107.4 tok/s85.9–128.8calibrated estimate ±20%	Q4_K_M	3.5 / 25.2 GB	Unified memory	up to 128k	thinking mode
Llama 3.2 3B	Runs great	est. 105.3 tok/s84.2–126.3calibrated estimate ±20%	Q4_K_M	3.5 / 25.2 GB	Unified memory	up to 64k	—
Qwen3.5 4B	Runs great	est. 100.6 tok/s80.5–120.7calibrated estimate ±20%	Q4_K_M	3.7 / 25.2 GB	Unified memory	up to 128k	thinking mode
MiniCPM-V 4.7 35B-A3B	Runs great	est. 99.7 tok/s79.8–119.7calibrated estimate ±20%	Q4_K_M	21.9 / 25.2 GB	Unified memory	up to 64k	thinking mode
Qwen3.6 35B-A3B	Runs great	est. 93.2 tok/s74.5–111.8calibrated estimate ±20%	Q4_K_M	23.0 / 25.2 GB	Unified memory	up to 64k	thinking mode
gpt-oss-20b	Runs great	92.4 tok/s8k estimate 101.6 tok/s (81.3–121.9, calibrated estimate ±20%)measured (1 run, 2k context)	MXFP4	12.9 / 25.2 GB	Unified memory	up to 128k	thinking mode
Kanana 1.5 15.7B-A3B	Runs great	est. 84.6 tok/s67.7–101.5calibrated estimate ±20%	Q4_K_M	12.1 / 25.2 GB	Unified memory	up to 32k	—
Gemma 4 E4B	Runs great	est. 83.8 tok/s58.6–108.9theoretical estimate ±30%	Q4_K_M	5.7 / 25.2 GB	Unified memory	up to 128k	thinking mode
Qwen3.5 35B-A3B	Runs great	est. 82.8 tok/s66.2–99.3calibrated estimate ±20%	Q4_K_M	23.4 / 25.2 GB	Unified memory	up to 64k	thinking mode
Qwen3 30B-A3B (2507)	Runs great	est. 80.0 tok/s64.0–96.0calibrated estimate ±20%	Q4_K_M	19.9 / 25.2 GB	Unified memory	up to 32k	—
Gemma 4 26B-A4B	Runs great	est. 71.2 tok/s56.9–85.4calibrated estimate ±20%	Q4_K_M	17.7 / 25.2 GB	Unified memory	up to 128k	thinking mode
DeepSeek R1 Distill Llama 8B	Runs great	est. 52.0 tok/s41.6–62.4calibrated estimate ±20%	Q4_K_M	6.6 / 25.2 GB	Unified memory	up to 32k	thinking mode
Kanana 1.5 8B	Runs great	est. 52.0 tok/s41.6–62.4calibrated estimate ±20%	Q4_K_M	6.6 / 25.2 GB	Unified memory	up to 32k	—
Llama 3.1 8B	Runs great	est. 52.0 tok/s41.6–62.4calibrated estimate ±20%	Q4_K_M	6.6 / 25.2 GB	Unified memory	up to 64k	—
K2-Horizon MoVA 36B-A4B	Runs great	est. 51.9 tok/s36.3–67.5theoretical estimate ±30%	Q4_K_M	24.6 / 25.2 GB	Unified memory	up to 8k	thinking mode
Nanbeige4.2 3B	Runs great	est. 50.9 tok/s35.6–66.1theoretical estimate ±30%	Q4_K_M	4.7 / 25.2 GB	Unified memory	up to 16k	thinking mode
Qwen3.5 9B	Runs great	est. 50.8 tok/s40.6–60.9calibrated estimate ±20%	Q4_K_M	6.7 / 25.2 GB	Unified memory	up to 128k	thinking mode
DeepSeek R1 0528 Qwen3 8B	Runs great	est. 50.0 tok/s40.0–59.9calibrated estimate ±20%	Q4_K_M	6.8 / 25.2 GB	Unified memory	up to 32k	thinking mode
Qwen3 8B	Runs great	est. 50.0 tok/s40.0–59.9calibrated estimate ±20%	Q4_K_M	6.8 / 25.2 GB	Unified memory	up to 32k	thinking mode
K2-Horizon 7B	Runs great	est. 45.8 tok/s36.7–55.0calibrated estimate ±20%	Q4_K_M	7.4 / 25.2 GB	Unified memory	up to 16k	thinking mode
Gemma 4 12B	Runs great	est. 43.0 tok/s34.4–51.5calibrated estimate ±20%	Q4_K_M	7.8 / 25.2 GB	Unified memory	up to 128k	thinking mode
Gemma 3 12B	Runs great	est. 39.8 tok/s31.8–47.7calibrated estimate ±20%	Q4_K_M	8.4 / 25.2 GB	Unified memory	up to 64k	—
HyperCLOVA X SEED Think 14B	Runs great	est. 30.6 tok/s21.4–39.7theoretical estimate ±30%	Q4_K_M	10.8 / 25.2 GB	Unified memory	up to 8k	thinking mode
Qwen3 14B	Runs great	est. 30.1 tok/s24.1–36.2calibrated estimate ±20%	Q4_K_M	10.9 / 25.2 GB	Unified memory	up to 8k	thinking mode
Phi-4	Runs great	est. 29.0 tok/s23.2–34.9calibrated estimate ±20%	Q4_K_M	11.3 / 25.2 GB	Unified memory	up to 16k	—
DeepSeek R1 Distill Qwen 14B	Runs well	est. 29.4 tok/s23.5–35.3calibrated estimate ±20%	Q4_K_M	11.2 / 25.2 GB	Unified memory	up to 64k	Try IQ4_XS: Runs great thinking mode
Mistral Small 3.2 24B	Runs well	est. 19.9 tok/s15.9–23.9calibrated estimate ±20%	Q4_K_M	16.2 / 25.2 GB	Unified memory	up to 32k	Try IQ4_XS: Runs great
Gemma 3 27B	Runs well	est. 18.1 tok/s14.5–21.7calibrated estimate ±20%	Q4_K_M	17.8 / 25.2 GB	Unified memory	up to 64k	Try IQ4_XS: Runs great
Qwen3.6 27B	Runs well	est. 18.0 tok/s14.4–21.5calibrated estimate ±20%	Q4_K_M	17.9 / 25.2 GB	Unified memory	up to 64k	thinking mode
Muse Glimmer 30B	Runs well	est. 17.9 tok/s14.3–21.5calibrated estimate ±20%	Q4_K_M	18.0 / 25.2 GB	Unified memory	up to 128k	thinking mode
Qwen3.5 27B	Runs well	est. 17.7 tok/s14.1–21.2calibrated estimate ±20%	Q4_K_M	18.2 / 25.2 GB	Unified memory	up to 64k	thinking mode
Qwen3.8 27B	Runs well	est. 17.3 tok/s13.9–20.8calibrated estimate ±20%	Q4_K_M	18.6 / 25.2 GB	Unified memory	up to 64k	thinking mode
Gemma 4 31B	Runs well	est. 16.4 tok/s13.1–19.7calibrated estimate ±20%	Q4_K_M	19.6 / 25.2 GB	Unified memory	up to 64k	thinking mode
Granite 4.2 30B	Runs well	est. 15.7 tok/s12.5–18.8calibrated estimate ±20%	Q4_K_M	20.4 / 25.2 GB	Unified memory	up to 16k	thinking mode
EXAONE 4.0 32B	Runs well	est. 15.7 tok/s12.5–18.8calibrated estimate ±20%	Q4_K_M	20.5 / 25.2 GB	Unified memory	up to 64k	thinking mode
EXAONE 4.5 33B	Runs well	est. 15.1 tok/s12.1–18.2calibrated estimate ±20%	Q4_K_M	21.2 / 25.2 GB	Unified memory	up to 32k	thinking mode
Qwen3 32B	Runs well	est. 14.2 tok/s11.4–17.1calibrated estimate ±20%	Q4_K_M	22.5 / 25.2 GB	Unified memory	up to 16k	thinking mode
DeepSeek R1 Distill Qwen 32B	Runs well	est. 14.2 tok/s11.3–17.0calibrated estimate ±20%	Q4_K_M	22.6 / 25.2 GB	Unified memory	up to 16k	thinking mode
Nanbeige4.3 33B	Runs well	est. 13.8 tok/s11.1–16.6calibrated estimate ±20%	Q4_K_M	23.1 / 25.2 GB	Unified memory	up to 16k	thinking mode
Nemotron 3.5 Lightning 30B-A3B	Runs slowly	est. 71.4 tok/s49.9–92.8theoretical estimate ±30%	Q4_K_M	25.5 GB RAM	CPU only	up to 256k	CPU inference — capped at Runs slowly Try IQ4_XS: Runs great thinking mode
Kolibri 1	Won't runTry Q2_K (heavy quality loss): Runs slowly	est. 85.3 tok/s59.7–110.9theoretical estimate ±30%	Q2_K	30.1 GB RAM	CPU only	up to 64k	CPU inference — capped at Runs slowly thinking mode
Llama 3.3 70B	Won't runTry Q2_K (heavy quality loss): Runs slowly	est. 7.1 tok/s4.9–9.2theoretical estimate ±30%	Q2_K	29.1 GB RAM	CPU only	up to 16k	—
DeepSeek V4 Flash 0731	Won't run	—	MXFP4	needs 155.6 GB	—	—	Needs about 155.6 GB; 32.0 GB of memory is free (25.2 GB usable by the GPU) thinking mode
GLM-5.3 Flash	Won't run	—	Q4_K_M	needs 182.1 GB	—	—	Needs about 182.1 GB; 32.0 GB of memory is free (25.2 GB usable by the GPU) thinking mode
gpt-oss-120b	Won't run	—	MXFP4	needs 64.3 GB	—	—	Needs about 64.3 GB; 32.0 GB of memory is free (25.2 GB usable by the GPU) thinking mode
Laguna S 2.1	Won't run	—	Q4_K_M	needs 68.6 GB	—	—	Needs about 68.6 GB; 32.0 GB of memory is free (25.2 GB usable by the GPU) thinking mode
Ling-3.0 Flash	Won't run	—	Q4_K_M	needs 77.7 GB	—	—	Needs about 77.7 GB; 32.0 GB of memory is free (25.2 GB usable by the GPU) thinking mode
Qwen3.5 122B-A10B	Won't run	—	Q4_K_M	needs 79.0 GB	—	—	Needs about 79.0 GB; 32.0 GB of memory is free (25.2 GB usable by the GPU) thinking mode
Solar Open 100B	Won't run	—	Q4_K_M	needs 64.4 GB	—	—	Needs about 64.4 GB; 32.0 GB of memory is free (25.2 GB usable by the GPU) thinking mode
Solar Open 2 250B	Won't run	—	IQ4_XS	needs 137.2 GB	—	—	Needs about 137.2 GB; 32.0 GB of memory is free (25.2 GB usable by the GPU) thinking mode

Models with a thinking mode spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s). With thinking turned off, a model may do one grade better.

Measured results on the Apple M4 Max (32-core GPU)

Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.

Measured results on the Apple M4 Max (32-core GPU)
Model	Quant	Backend	Context	Prompt (tok/s)	Generation (tok/s)	Flags	Source	Measured
gpt-oss-20b	MXFP4	llama.cpp	2k	1,277.0	92.4	—	github.com	2025-08-18

Frequently asked questions

What is the largest model that runs entirely on the Apple M4 Max (32-core GPU)?

K2-Horizon MoVA 36B-A4B at Q4_K_M (a 22.37 GB file) fits entirely in 36 GB of unified memory with 8k context, at est. 51.9 tok/s (36.3–67.5, theoretical estimate ±30%).

How many local LLMs run well on the Apple M4 Max (32-core GPU)?

At Q4_K_M with 8k context, with 36 GB of memory, out of 59 tracked models: 34 run great, 14 run well, 1 runs slowly and 10 won't run.

Can the Apple M4 Max (32-core GPU) run a 70B model like Llama 3.3 70B?

Runs slowly — CPU only, Q2_K (heavy quality loss): est. 7.1 tok/s (4.9–9.2, theoretical estimate ±30%).

Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.