FitMyLLM
APPLE· M3 Max

Apple M3 Max (48GB)

Running LLMs on the Apple M3 Max (48GB) — the long read: which models fit at which quantisation, and the settings worth changing. · Or what else $2,899 buys

Usable memory
36 GB
High-end
Bandwidth
400
GB/s
Models Q4
367/498
74%
7B Q4 speed
~51
Blazing
Model coverage at Q474% of all
Estimated speed· By model size at Q4

Average speeds at Q4 quantization. Actual performance varies by model architecture and context length.

3B
~119
tok/s
7B
~51
tok/s
14B
~25
tok/s
32B
~11
tok/s
70B
—
39.4GB needed
Measured rig reports

We rent the machine and time every model on it: decode, VRAM peak, concurrency, watts and cost per million tokens. Yours may already be one of them — and two are free to read in full.

See the reports →
Specifications
Usable memory
36 GB of 48
Bandwidth
400 GB/s
FP16 Compute
16.4 TFLOPS
TDP
60W
Memory
Unified
Architecture
M3 Max
MSRP
$2,899
272
Fast models · >30 tok/s
Real-time chat speed
364
Usable · >10 tok/s
Comfortable for all tasks
367
Total compatible
Fit in usable memory at Q4
Don’t want to buy?

Test M3 Max (48GB) (or anything bigger) without committing. Pay by the second, cancel anytime.

Spin up in ~60s. Pay by the second. Cancel anytime.

Some links are affiliate links — we may earn a small commission at no extra cost to you. This helps keep FitMyLLM free and independent.

Compatible models· 367
S
Falcon-H1R Tiny 90M0.09B
Falcon·256K ctx· chat· reasoning
3951
tok/s · 2% mem
›
S
SmolLM2 135M0.135B
Smollm·2K ctx· chat
2634
tok/s · 2% mem
›
S
nomic-embed-text-v1.5 100M0.14B
Embedding·8K ctx· chat
2540
tok/s · 2% mem
›
S
GPT-2 124M0.14B
Gpt2·1K ctx· chat
2540
tok/s · 2% mem
›
S
SmolVLM 256M0.256B
Smollm·8K ctx· chat· vision
1389
tok/s · 2% mem
›
S
Gemma 3 270M0.27B
Gemma·32K ctx· chat
1317
tok/s · 2% mem
›
S
Snowflake Arctic Embed M v2.00.305B
Embedding·8K ctx· embedding· multilingual
1166
tok/s · 2% mem
›
S
bge-large-en-v1.5 335M0.335B
Embedding·1K ctx· chat
1061
tok/s · 2% mem
›
S
mxbai-embed-large-v10.335B
Embedding·1K ctx· embedding
1061
tok/s · 2% mem
›
S
Snowflake Arctic Embed L0.335B
Embedding·1K ctx· embedding
1061
tok/s · 2% mem
›
S
LFM2 350M0.35B
Lfm·125K ctx· chat· tool_use
1016
tok/s · 2% mem
›
S
SmolLM2 360M0.36B
Smollm·8K ctx· chat
988
tok/s · 2% mem
›
S
GPT-2 Medium 345M0.38B
Gpt2·1K ctx· chat
936
tok/s · 2% mem
›
S
Qwen 2.5 0.5B0.5B
Qwen·32K ctx· chat
711
tok/s · 2% mem
›
S
SmolVLM 500M0.5B
Smollm·8K ctx· chat· vision
711
tok/s · 2% mem
›
S
Falcon-H1 0.5B0.52B
Falcon·128K ctx· chat
684
tok/s · 2% mem
›
S
BGE-M30.568B
Embedding·8K ctx· embedding
626
tok/s · 2% mem
›
S
Snowflake Arctic Embed L v2.00.568B
Embedding·8K ctx· embedding· multilingual
626
tok/s · 2% mem
›
S
Qwen3-Embedding 0.6B0.6B
Embedding·32K ctx· embedding· multilingual
593
tok/s · 2% mem
›
S
Falcon-H1R Tiny 0.6B0.6B
Falcon·32K ctx· chat· reasoning
593
tok/s · 2% mem
›
S
Falcon Perception 0.6B0.6B
Falcon·4K ctx· vision
593
tok/s · 2% mem
›
S
Qwen 1.5 0.5B0.62B
Qwen·32K ctx· chat
573
tok/s · 2% mem
›
S
LFM2 700M0.74B
Lfm·125K ctx· chat· tool_use
480
tok/s · 3% mem
›
S
Qwen3 0.6B0.75B
Qwen·32K ctx· chat· reasoning
474
tok/s · 3% mem
›
S
GPT-2 Large 774M0.81B
Gpt2·1K ctx· chat
439
tok/s · 3% mem
›
S
Qwen 3.5 0.8B0.87B
Qwen·256K ctx· chat· coding· multilingual
409
tok/s · 3% mem
›
S
InternVL3 1B0.94B
Other·32K ctx· chat· vision
378
tok/s · 3% mem
›
S
MiniCPM5 1B1.08B
Minicpm·128K ctx· chat· reasoning· multilingual
329
tok/s · 3% mem
›
S
K2 Horizon 0.9B1.08B
Other·128K ctx· chat· coding· reasoning
329
tok/s · 3% mem
›
S
TinyLlama 1.1B1.1B
Llama·2K ctx· chat
323
tok/s · 3% mem
›
S
LFM2.5-1.2B-Thinking1.2B
Lfm·122K ctx· chat· reasoning· tool_use
296
tok/s · 3% mem
›
S
Llama-3.2-1B1.2B
Llama·4K ctx· chat
296
tok/s · 3% mem
›
S
LFM2 1.2B1.2B
Lfm·125K ctx· chat· tool_use· multilingual
296
tok/s · 3% mem
›
S
Zamba2 1.2B1.2B
Other·4K ctx· chat
296
tok/s · 3% mem
›
S
Ling 3.0 Tiny7.89BMoE
Other·128K ctx· chat· coding· reasoning
274
tok/s · 15% mem
›
S
EXAONE-4.0-1.2B1.3B
Exaone·64K ctx· chat
274
tok/s · 4% mem
›
S
OPT 1.3B1.3B
Opt·2K ctx· chat
274
tok/s · 4% mem
›
S
MiniCPM-V 4.61.3B
Other·256K ctx· chat· vision
274
tok/s · 4% mem
›
S
DeepSeek Coder 1.3B1.35B
Deepseek·16K ctx· coding
263
tok/s · 4% mem
›
S
Phi-1 1.3B1.42B
Phi·2K ctx· coding
250
tok/s · 4% mem
›
Next step

Get personalized recommendations.

See ranked models with benchmark scores, run commands, and precise speed estimates for your Apple M3 Max (48GB).

What this card is worth

M3 Max (48GB) holds 367 of the models in our catalogue and is, in practice, a IQ4_XS card — the largest it takes is Jamba 2 Mini at IQ4_XS.

Tokens/sec per $100
1.6
8B at Q4_K_M, so cards compare like for like
Usable memory per $100
1.24 GB
what memory costs on this card
The biggest it takes
Jamba 2 Mini52B · IQ4_XS31.6 GB30 tok/sest
Jamba 1.5 Mini 52B51.6B · IQ4_XS31.3 GB30 tok/sest
Kimi-Linear-48B-A3B48B · Q4_K_M32.0 GB110 tok/sest
Nemotron-H 47B47B · Q4_K_M31.3 GB7 tok/sest
Mixtral-8x7B46.7B · Q4_K_M31.5 GB25 tok/sest
Nous-Hermes-2-Mixtral-8x7B-DPO46.7B · Q4_K_M31.5 GB25 tok/sest
Dolphin 2.6 Mixtral 8x7B46.7B · Q4_K_M31.5 GB25 tok/sest
Phi-3.5 MoE 42B41.9B · Q5_K_S32.2 GB44 tok/sest
Against cards of similar memory
M1 Ultra (64GB)48 GB$2,4992.9 tok/s per $100
M2 Ultra (64GB)48 GB$2,9992.4 tok/s per $100
M4 Max (64GB)48 GB$2,8991.9 tok/s per $100
M2 Max (64GB)48 GB$2,2992.0 tok/s per $100
M3 Max (48GB)36 GB$2,8991.6 tok/s per $100

Neighbours in memory rather than in price: memory decides whether a card can do the job at all, so two cards of the same size at different prices is the comparison you are making. Ordered by memory, then bandwidth — not by the value column, which is worked out from bandwidth and price alone and therefore rewards a cheap card whatever its software stack does to that bandwidth in practice. Read it as one input, not as a ranking.

Device under test

Apple M3 Max (48GB) — 48 GB unified, 36 GB usable.

Apple M3 Max (48GB) spec
Brand
Apple
Unified memory
48 GB
Usable by a model
36 GB
Bandwidth
400 GB/s
FP16 Compute
16.4 TFLOPS
FP32 Compute
16.4 TFLOPS
TDP
60 W
Architecture
M3 Max
MSRP
$2899
AI capability
367/ 498 models @ Q4

With 36 GB of its 48 GB reaching a model and 400 GB/s bandwidth, this machine handles models up to 48B parameters.

Speed ≈ bandwidth / model_size × efficiency. A 7B model at Q4 runs at ~46 tok/s.

Rent it — live prices
Checking what the clouds are charging…
Top models for Apple M3 Max (48GB)
367 fit · showing 20
ModelSizeVRAM Q4tok/sAvg
Kimi-Linear-48B-A3B48B29.8 GB11926.6
Nemotron-H 47B47B29.2 GB873.8
Mixtral-8x7B46.7B29.0 GB2718.8
Nous-Hermes-2-Mixtral-8x7B-DPO46.7B29.0 GB2727.4
Dolphin 2.6 Mixtral 8x7B46.7B29.0 GB2723.8
Phi-3.5 MoE 42B41.9B26.1 GB5456.7
Falcon 40B40B24.9 GB920.9
Flint39.79B24.8 GB14277.3
InternVL3 38B38B23.7 GB978.9
K2 Horizon MoVA 36B-A4B37.44B23.4 GB8945.7
Seed-OSS 36B Instruct36B22.5 GB1054.4
MiniCPM-V 4.7 35B-A3B35.21B22.0 GB11976.4
c4ai-command-r-v01 35B35B21.9 GB1027.5
Qwen 3.5 35B A3B35B21.9 GB11956.1
Qwen 3.6 35B A3B35B21.9 GB11953.3
K2 Horizon 32B34.78B21.7 GB1054.0
Nous Capybara 34B34.4B21.5 GB1042.0
Yi-1.5 34B34.4B21.5 GB1045.3
Falcon-H1 34B34B21.3 GB1058.6
CodeLlama 34B34B21.3 GB1025.4