LocalLLMChecker

Unified memory · Apple

What AI models can an Apple M2 Max 96GB run?

96 GB is server territory. The constraint stops being memory and becomes bandwidth: everything fits, and what varies is how long you wait for it. Memory is shared with the CPU, so the practical ceiling is lower than the sticker number: macOS hands roughly three quarters of it to the GPU by default.

Memory96 GB
Bandwidth400 GB/s
Usable for a model72 GB
Runtime backendMETAL

The short answer

Assuming an 8k context window and default settings, these are the models worth downloading first.

Best all-rounder

GPT OSS 120B

MXFP4 · 59.65 GB · about 50 tokens/s

Best for code

Qwen3 32B

Q6_K · 27.43 GB · about 10.3 tokens/s

Largest that still runs well

GPT OSS 120B

116.83B parameters · MXFP4 · about 50 tokens/s

Fastest useful answer

LFM2.5 8B A1B

about 95.7 tokens/s · 8.78 GB

See how fast it feels

Apple M2 Max 96GB running GPT OSS 120B at MXFP4

Wait for the first word1.4 s
Then writes at50 tok/s
Whole answer4.9 s

YouWhy does my model use more memory when the conversation gets longer?

Model

Simulated from our estimate at a 512-token question, not a recording. Assumes nothing else is competing for the GPU.

Where it sits

The GPU can use about 72 GB of this machine's 96 GB by default, so for planning it behaves like a 48 GB card with different bandwidth. The 48 GB guide explains what that tier runs and where it runs out, and the bandwidth guide explains why 400 GB/s decides how fast it answers.

Our indicative price for this machine is $3,699 new, reviewed 2026-09-01. That is 2.6 GB of memory and 11 GB/s of bandwidth for every $100. Prices move, so treat these as a place to start a search and not a quote.

Devices in a similar price band, fastest memory first:

DevicePriceMemoryBandwidthBest all-round pick
Apple M3 Ultra 96GB $3,999 96 GB 819 GB/s GPT OSS 120B, 102.5 t/s
Apple M1 Ultra 128GB $3,400 128 GB 800 GB/s GPT OSS 120B, 100.1 t/s
Apple M2 Ultra 64GB $3,499 64 GB 800 GB/s Qwen3.5 27B, 20.2 t/s
Apple M4 Max 36GB $3,199 36 GB 546 GB/s Gemma 4 12B Instruct, 28.2 t/s

Every model, scored on this device

Each row uses the highest-quality quantization that both fits and stays conversational. Speed is a single-stream estimate at 8k context.

Model Params Verdict Download Memory used Speed Max context
GPT OSS 120B
GPT-OSS 		116.83B (5.7B active) Runs great MXFP4 59.65 GB 50 t/s 128k
Qwen3.5 35B A3B
Qwen 		35.95B (2.9B active) Runs great Q8_0 36.1 GB 51.8 t/s 256k
Gemma 4 26B A4B Instruct
Gemma 		25.81B (3.74B active) Runs great Q8_0 26.69 GB 34.9 t/s 256k
GLM 4.7 Flash
GLM 		31.22B (3.66B active) Runs great Q8_0 31.61 GB 39.5 t/s 128k
Gemma 4 31B Instruct
Gemma 		31.27B Runs well Q6_K 27.58 GB 10 t/s 128k
Qwen3 30B A3B
Qwen 		30.53B (3.34B active) Runs great Q8_0 31.27 GB 39.2 t/s 32k
Phi 3.5 MoE Instruct
Phi 		41.87B (6.64B active) Runs well Q8_0 42.81 GB 20.9 t/s 128k
Gemma 4 12B Instruct
Gemma 		11.96B Runs well Q8_0 13.52 GB 20.7 t/s 256k
Qwen3 32B
Qwen 		32.76B Runs well Q6_K 27.43 GB 10.3 t/s 32k
GPT OSS 20B
GPT-OSS 		20.91B (4.18B active) Runs great MXFP4 11.81 GB 66 t/s 128k
Qwen3.5 27B
Qwen 		27.78B Runs well Q8_0 28.54 GB 10.1 t/s 256k
Qwen3 Coder 30B A3B Instruct
Qwen 		30.53B (3.34B active) Runs great Q8_0 31.27 GB 39.2 t/s 32k
Qwen3 14B
Qwen 		14.77B Runs well Q8_0 16.28 GB 18 t/s 32k
Granite 4.1 30B
Granite 		28.87B Runs well Q6_K 24.42 GB 11.3 t/s 128k
Qwen3.5 9B
Qwen 		9.65B Runs great Q8_0 10.22 GB 30.8 t/s 256k
Qwen2.5 14B Instruct
Qwen 		14.77B Runs well Q8_0 16.53 GB 17.7 t/s 32k
Qwen2.5 32B Instruct
Qwen 		32.76B Runs well Q6_K 27.43 GB 10.3 t/s 32k
Gemma 3 27B Instruct
Gemma 		27.43B Runs well Q6_K 22.45 GB 12.3 t/s 128k
Qwen2.5 Coder 14B Instruct
Qwen 		14.77B Runs well Q8_0 16.53 GB 17.7 t/s 32k
Qwen2.5 Coder 32B Instruct
Qwen 		32.76B Runs well Q6_K 27.43 GB 10.3 t/s 32k
Gemma 3 12B Instruct
Gemma 		12.19B Runs well Q8_0 13.24 GB 21.1 t/s 128k
Phi 4
Phi 		14.66B Runs well Q8_0 16.48 GB 17.5 t/s 16k
Gemma 4 E4B Instruct
Gemma 		8B Runs great Q8_0 8.39 GB 33.7 t/s 128k
Mixtral 8x7B Instruct V0.1
Mistral 		46.7B (12.88B active) Runs well Q8_0 47.59 GB 11.5 t/s 32k
Granite 4.1 8B
Granite 		8.79B Runs great Q8_0 10.32 GB 27.3 t/s 128k
Qwen3 8B
Qwen 		8.19B Runs great Q8_0 9.61 GB 31.6 t/s 32k
Mistral Nemo Instruct 2407
Mistral 		12.25B Runs well Q8_0 13.78 GB 21.4 t/s 128k
Qwen2.5 VL 7B Instruct
Qwen 		8.29B Runs great Q8_0 9 GB 33.6 t/s 64k
Mistral Small 24B Instruct 2501
Mistral 		23.57B Runs well Q8_0 24.98 GB 11.4 t/s 32k
Qwen2.5 7B Instruct
Qwen 		7.62B Runs great Q8_0 8.34 GB 36.6 t/s 32k
Show the other 56 models
Llama 3.1 8B Instruct 8.03B Runs great Q8_0 9.32 GB 32.3 t/s 128k
DeepSeek Coder V2 Lite Instruct 15.71B Runs great Q8_0 16.09 GB 54.2 t/s 128k
Codestral 22B V0.1 22.25B Runs well Q8_0 24.21 GB 11.5 t/s 32k
Gemma 2 9B Instruct 9.24B Runs well Q8_0 11.47 GB 24.5 t/s 8k
Qwen2.5 Coder 7B Instruct 7.62B Runs great Q8_0 8.34 GB 36.6 t/s 32k
OLMo 3 7B Instruct 7.3B Runs great Q8_0 10.1 GB 29.2 t/s 64k
Internlm3 8B Instruct 8.8B Runs great Q8_0 9.46 GB 31.8 t/s 32k
Mistral 7B Instruct V0.3 7.25B Runs great Q8_0 8.55 GB 33.8 t/s 32k
Gemma 4 E2B Instruct 5.12B Runs great Q8_0 5.41 GB 53.1 t/s 128k
Qwen3.5 4B 4.66B Runs great Q8_0 5.24 GB 55.4 t/s 256k
Mistral 7B Instruct V0.2 7.24B Runs great Q8_0 8.54 GB 33.8 t/s 32k
Llama 3 3 Nemotron Super 49B V1 49.87B Fits, but slow Q8_0 69.86 GB 4 t/s 8k
DeepSeek Coder 7B Instruct V1.5 6.91B Runs great Q8_0 10.96 GB 26.7 t/s 4k
DeepSeek Coder 6.7B Instruct 6.74B Runs great Q8_0 11.04 GB 25.8 t/s 16k
LFM2.5 8B A1B 8.47B Runs great Q8_0 8.78 GB 95.7 t/s 64k
Qwen3 4B 4.02B Runs great Q8_0 5.43 GB 53.3 t/s 32k
Gemma 3 4B Instruct 4.3B Runs great Q8_0 4.88 GB 59.7 t/s 128k
Phi 4 Mini Instruct 3.84B Runs great Q8_0 5.14 GB 56.7 t/s 128k
Granite 4.1 3B 3.4B Runs great Q8_0 4.31 GB 68.2 t/s 128k
Qwen2.5 3B Instruct 3.09B Runs great Q8_0 3.65 GB 81.4 t/s 32k
Phi 3 Mini 4k Instruct 3.82B Runs great Q8_0 4.87 GB 61.4 t/s 4k
Llama 3.2 3B Instruct 3.21B Runs great Q8_0 4.39 GB 67.1 t/s 128k
LFM2.5 2.6B 2.7B Runs great Q8_0 3.11 GB 97.2 t/s 128k
Qwen3.5 2B 2.27B Runs great Q8_0 2.67 GB 115.3 t/s 256k
Gemma 2 2B Instruct 2.61B Runs great Q8_0 3.51 GB 85.2 t/s 8k
Starcoder2 3B 3.03B Runs great Q8_0 3.46 GB 91.7 t/s 16k
Qwen3 1.7B 2.03B Runs great Q8_0 3.19 GB 94.3 t/s 32k
DeepSeek R1 Distill Qwen 1.5B 1.78B Runs great Q8_0 2.27 GB 155.5 t/s 128k
Qwen2.5 1.5B Instruct 1.54B Runs great Q8_0 2.03 GB 156.1 t/s 32k
SmolLM2 1.7B 1.71B Runs great Q8_0 3.5 GB 85.2 t/s 8k
OLMo 2 0425 1B 1.48B Runs great Q8_0 2.77 GB 120.3 t/s 4k
Llama 3.2 1B Instruct 1.24B Runs great Q8_0 1.79 GB 184.1 t/s 128k
LFM2.5 1.2B Instruct 1.17B Runs great Q8_0 1.56 GB 243.1 t/s 64k
Gemma 3 1B Instruct 1B Runs great Q8_0 1.32 GB 261.8 t/s 32k
Qwen3.5 0.8B 0.87B Runs great Q8_0 1.25 GB 279.5 t/s 256k
Qwen3 0.6B 0.75B Runs great Q8_0 1.89 GB 168.2 t/s 32k
Qwen2.5 0.5B Instruct 0.49B Runs great Q8_0 0.85 GB 470 t/s 32k
LFM2.5 350M 0.35B Runs great Q8_0 0.71 GB 727.8 t/s 64k
Qwen3 Next 80B A3B Instruct 81.32B Just fits n/a 81.04 GB 46.3 t/s n/a
Qwen3 Coder Next 79.67B Just fits n/a 79.4 GB 46.3 t/s n/a
Qwen3.5 122B A10B 125.09B Just fits n/a 83.53 GB 27.6 t/s n/a
MiniMax M2.5 228.7B Fits only at 2 to 3 bits n/a 83.75 GB 27 t/s n/a
MiniMax M2.7 228.69B Fits only at 2 to 3 bits n/a 83.75 GB 27 t/s n/a
Qwen3 235B A22B 235.09B Fits only at 2 to 3 bits n/a 85.59 GB 16.9 t/s n/a
GLM 4.5 Air 110.47B Just fits n/a 86.18 GB 13.6 t/s n/a
Llama 3.3 70B Instruct 70.55B Just fits n/a 72.82 GB 3.8 t/s n/a
Qwen2.5 VL 72B Instruct 73.41B Just fits n/a 75.65 GB 3.7 t/s n/a
Qwen2.5 72B Instruct 72.71B Just fits n/a 74.96 GB 3.7 t/s n/a
Kimi K2.6 1026.88B Will not run n/a 578.41 GB n/a n/a
Kimi K2 Instruct 1026.41B Will not run n/a 578.15 GB n/a n/a
GLM 5.1 753.86B Will not run n/a 425.01 GB n/a n/a
GLM 5.2 753.33B Will not run n/a 424.72 GB n/a n/a
DeepSeek V3.2 685.4B Will not run n/a 386.4 GB n/a n/a
DeepSeek R1 684.53B Will not run n/a 385.92 GB n/a n/a
MiniMax M3 427.04B Will not run n/a 241.5 GB n/a n/a
GLM 4.5 358.34B Will not run n/a 204.77 GB n/a n/a

What it will not run

18 of the 86 architectures we track are out of reach here, even at 2-bit precision.

If you want more headroom

The next steps up in memory, in order. More memory changes which models load at all; more bandwidth changes how fast they answer, so the two are worth weighing separately.

DeviceMemoryBandwidthModels that fit
NVIDIA DGX Spark 128GB 128 GB 273 GB/s 75 Check price
Ryzen AI Max+ 395 128GB 128 GB 256 GB/s 75 Check price

Price links go to an Amazon search for the model name and are affiliate links: if you buy through one, we earn a commission at no cost to you. We do not take payment for placement, and the ordering above is by memory capacity alone.

Other devices with 96 GB

Same capacity, different speed. Once a model fits, bandwidth is what separates these.

RTX PRO 6000 Blackwell

1792 GB/s · NVIDIA

Apple M3 Ultra 96GB

819 GB/s · Apple

Apple M3 Max 96GB

400 GB/s · Apple

Ryzen AI Max+ 395 96GB

256 GB/s · AMD

Questions people ask about an Apple M2 Max 96GB

How much VRAM does an Apple M2 Max 96GB have?

The Apple M2 Max 96GB has 96 GB of unified memory shared with the CPU, of which roughly 72 GB is available to a model once the operating system takes its share.

Can an Apple M2 Max 96GB run a 70B model?

Yes, once the macOS GPU memory limit is raised. At Q8_0 it needs 79.4 GB and runs at around 46.3 tokens per second.

What is the best local model for an Apple M2 Max 96GB?

GPT OSS 120B at MXFP4 is the strongest all-round fit: 59.65 GB used, around 50 tokens per second at 8k context.