Local LLMs on a Mac
Scenario A
Add a scenario
CurrencySaveShare
Parameters
Mac
Mac
Chip
Unified memory
24 GB48 GB64 GB
Model
Model
Quantization
3-bit4-bit6-bit8-bitOriginal
Usage
Context
tokens
4,096 tokens16,384 tokens32,768 tokens131,072 tokens
More settings
Parameters
Mac
MacMacBook Pro 14-inch
ChipM5 Pro, 18-core CPU, 20-core GPU
Unified memory48 GB
Model
ModelGemma 4 26B A4B
Quantization4-bit
Usage
Context32,768 tokens
RuntimeMLX
Raise the GPU memory limitNo
Gemma 4 26B A4B on the MacBook Pro 14-inch with 48 GB writes
83 tok/s
Much faster than you read. At the end of a 32,768-token context: 54 tok/s.
Reading speed
2,302 tok/s
Prompt processing, with the GPU's Neural Accelerators.
Time to read the context
17 s
A 32,768-token prompt, before the first word of the answer.
Writing at full context
54 tok/s
Once 32,768 tokens are in memory.
Memory needed
17 GB
40 GB usable by the GPU, out of 48 GB.
Memory
17 GB of 40 GB
- Weights (4.5 bits per weight)15 GB
- KV cache (32,768 tokens)0.9 GB
- Runtime buffers1.1 GB
23 GB left for the GPU.
M5 Pro, 18-core CPU, 20-core GPU
CPU18 cores (6 super + 12 performance)
GPU20 cores
Memory bandwidth307 GB/s
GPU compute (FP32)8.3 TFLOPS
Neural Accelerators (FP16)33 TFLOPS
Neural Engine16 cores
Price$3,299
Price: Apple Store United States, before sales tax, base storage. The currency picks the store.
The cheapest Macs for Gemma 4 26B A4B
In 4-bit, with a 32,768-token context. Prices: Apple Store United States, before sales tax.
Cheapest fast one (30 tok/s)
$1,099
Mac mini, M6 · 12 GPU, 24 GB
53 tok/s
Models on this Mac
23 of 39 models fit, in 4-bit, with a 32,768-token context (or the model's maximum when lower).
| Model | Memory | Writing | Verdict | | ----------------------------------------------------------- | ------ | --------- | ----------- | | Llama 3.2 1BMeta · 1.2B | 2.8 GB | 309 tok/s | Fast | | Qwen3.5 2BAlibaba · 2.3B | 3.2 GB | 181 tok/s | Fast | | Gemma 4 E2BGoogle · 5.1B | 4.9 GB | 179 tok/s | Fast | | Llama 3.2 3BMeta · 3.2B | 6.6 GB | 132 tok/s | Fast | | SmolLM3 3BHugging Face · 3.1B | 5.2 GB | 137 tok/s | Fast | | Nemotron 3 Nano 4BNVIDIA · 4B | 3.9 GB | 108 tok/s | Fast | | Qwen3.5 4BAlibaba · 4.7B | 5.2 GB | 93 tok/s | Fast | | Gemma 4 E4BGoogle · 8B | 6.7 GB | 96 tok/s | Fast | | Llama 3.1 8BMeta · 8B | 9.9 GB | 55 tok/s | Fast | | Granite 4.2 8BIBM · 8.8B | 11 GB | 50 tok/s | Fast | | Qwen3.5 9BAlibaba · 9.7B | 8.2 GB | 46 tok/s | Fast | | Gemma 4 12BGoogle · 12B | 8.6 GB | 37 tok/s | Fast | | Ministral 3 14BMistral AI · 14B | 15 GB | 32 tok/s | Fast | | Phi-4Microsoft · 14.7B | 13 GB | 30 tok/s | Fast | | gpt-oss 20BOpenAI · 20.9B, 3.6B active (MoE) | 14 GB | 77 tok/s | Fast | | Devstral Small 2 24BMistral AI · 24B | 22 GB | 19 tok/s | Comfortable | | Gemma 4 26B A4BGoogle · 25.2B, 3.8B active (MoE) | 17 GB | 83 tok/s | Fast | | Qwen3.8 27BAlibaba · 27.8B | 19 GB | 16 tok/s | Comfortable | | Muse Glimmer 30BMeta · 29.6B | 21 GB | 16 tok/s | Comfortable | | GLM-4.7-FlashZ.ai · 30B, 3B active (MoE) | 20 GB | 54 tok/s | Fast | | Nemotron 3.5 Lightning 30B A3BNVIDIA · 30B, 3B active (MoE) | 19 GB | 78 tok/s | Fast | | Gemma 4 31BGoogle · 30.7B | 23 GB | 15 tok/s | Comfortable | | Qwen3.6 35B A3BAlibaba · 35B, 3B active (MoE) | 22 GB | 88 tok/s | Fast |
Too big for this Mac: Llama 3.3 70B (52 GB), Qwen3-Coder-Next (47 GB), gpt-oss 120B (66 GB), Mistral Small 4 119B (70 GB), Nemotron 3 Super 120B A12B (70 GB), Qwen3.5 122B A10B (72 GB), Mistral Medium 3.5 128B (91 GB), Qwen3.8 Flash Next (114 GB), MiniMax M2.7 (138 GB), DeepSeek V4 Flash (153 GB), GLM-5.3-Flash (206 GB), Qwen3.5 397B A17B (226 GB), MiniMax M3 (249 GB), DeepSeek V3.2 (382 GB), GLM-5.3 (422 GB), Kimi K2.6 (581 GB).
Writing speed83 tok/s
How it works
Writing speed is set by the memory bandwidth
To write each token, the Mac reads every active weight of the model once. So the writing speed (decode) is about the memory bandwidth divided by the size of those weights. An M5 Max reads 614 GB per second: a dense 70B model at 4 bits weighs about 40 GB, so it writes about 13 tokens per second.
The simulator uses 75 to 87% of the bandwidth with MLX depending on the quantization (5 points less with llama.cpp), plus a fixed cost per token measured on each kind of chip and model. A long conversation slows things down: every new token also reads the KV cache, the model's memory of the context.
Reading speed is set by the GPU
Before answering, the model reads your prompt (prefill). That step is limited by compute: about 2 operations per active parameter and per token, plus attention, which grows with the square of the prompt length.
M5 and M6 chips have a Neural Accelerator in each GPU core. MLX uses them (macOS 26.2 and later): prompts are read 3.5 to 4 times faster than on M4. llama.cpp uses them since April 2026 for matrix products, but not yet for attention.
What decides if a model fits
The weights, the KV cache and the runtime's buffers must fit in the memory macOS lets the GPU use. On macOS 26 and 27 it is 74% of 16 or 24 GB, 78% of 32 to 48 GB, 81% of 64 or 96 GB, 84% of 128 GB, 87% of 256 GB and 464 GB of 512 GB. The command sudo sysctl iogpu.wired\_limit\_mb raises it until the next restart.
Apple's memory sizes are binary: 16 GB is 17.2 billion bytes. Model sizes here are in billions of bytes, like on Hugging Face.
Quantization and mixtures of experts
Quantization stores each weight in fewer bits. 4-bit is the usual choice: the model is 3.5 times smaller than in 16 bits and loses little quality. 8-bit is almost lossless. A bigger model in 4-bit usually beats a smaller one in 8-bit.
A mixture of experts (MoE) like gpt-oss, Qwen3.6 35B A3B or DeepSeek only uses a few experts per token. All of its weights must fit in memory, but each token only reads the active ones: gpt-oss 120B needs 65 GB and writes like a 5B model.
How reliable the numbers are
The constants come from about 300 published measurements: the llama.cpp benchmark thread, MLX and oMLX results, Apple's own figures and reviews of the M5 and M6 Macs. The estimates are usually within 15% of a measurement on the same software. Your speed depends on the app, its version and its settings: Ollama, LM Studio and llama.cpp do not all use the same engine.
Prices are those of the Apple Store of the chosen currency on October 1st, 2026, with base storage. Apple raised them on June 25th, 2026. French MacBooks, and British MacBook Air and Pro, come without a power adapter. The 512 GB Mac Studio ships in late October 2026: its price is not published yet.
Frequently asked questions
Which Mac do I need for gpt-oss 120B?
More than 64 GB of memory: it needs about 66 GB and 64 GB Macs let the GPU use 56 GB. The cheapest is the Mac Studio M5 Max with 128 GB ($5,099), at about 89 tokens per second. The MacBook Pro M5 Max with 128 GB is as fast. The Mac Studio M5 Ultra with 96 GB ($5,499) writes about 113.
How much memory does a 70B model need?
Llama 3.3 70B in 4-bit needs about 40 GB of weights, plus 11 GB of KV cache for a 32,768-token context. A 64 GB Mac runs it, slowly: about 6 tokens per second on an M5 Pro, 13 on an M5 Max. The M5 Ultra (96 GB at least) writes about 23.
Can a MacBook Neo or a MacBook Air run a local LLM?
Yes, small ones. The MacBook Neo has 8 GB: models up to 4B parameters, like Qwen3.5 4B at about 19 tokens per second. A 16 GB MacBook Air runs models up to 14B (Phi-4 at about 15 tokens per second, Llama 3.1 8B at about 28). They have no fan: long prompts slow down as the chip heats up.
MLX or llama.cpp (Ollama, LM Studio)?
On a Mac, MLX is usually a little faster. It is much faster on hybrid models like Qwen3.5 and Qwen3.6 (1.4 to 1.9 times), and it reads prompts faster on M5 and M6. LM Studio runs both. Ollama runs some models on MLX and the others on llama.cpp.
Why does a 35B model run faster than an 8B one?
Because it is a mixture of experts. Qwen3.6 35B A3B only uses 3 billion parameters per token: with MLX it writes faster than a dense 8B model, but its 35 billion parameters must all fit in memory.
Can I run DeepSeek or Kimi on a Mac?
DeepSeek V4 Flash, yes: in 4-bit it needs about 150 GB and runs on a 256 GB Mac Studio M5 Ultra, at about 48 tokens per second. DeepSeek V3.2 and Kimi need the 512 GB Mac Studio M5 Ultra, which ships in late October 2026. DeepSeek V3.2 in 4-bit needs about 380 GB and writes about 25 tokens per second. Kimi K2.6 (1,000 billion parameters) only fits in 3-bit.
Are my settings sent anywhere?
No. The math runs in your browser. Nothing is sent until you click Save or Share.
Other simulators
Rent or buy?Compare your net worth in both cases, year after year.Open the simulatorCompound interestWatch your savings grow with regular contributions.Open the simulatorSLA calculatorHow much downtime does your SLA really allow?Open the simulator