Apple Silicon for LLMs: M1 to M6 Complete Guide (2026)

Why Apple Silicon Is Different: Unified Memory Architecture

Apple Silicon's defining advantage for LLM inference is unified memory. On a discrete GPU system, the GPU can only access its dedicated VRAM. The 24 GB on an RTX 4090 is a hard ceiling, the system's 64 GB of DDR5 RAM is irrelevant. On Apple Silicon, the CPU and GPU share the same physical memory pool. This means an M4 Max with 128 GB of unified memory can hold a 128 GB model entirely in GPU-accessible memory, and the new M5 Ultra pushes that ceiling to 512 GB. This is how a MacBook Pro runs Llama 3.3 70B at Q8_0 (~72 GB) while an RTX 4090 desktop cannot even load it at Q4. The trade-off is memory bandwidth. Apple Silicon's bandwidth is shared between CPU and GPU, and it is lower than high-end discrete GPUs. The M4 Max delivers 546 GB/s, excellent for a laptop but roughly half the RTX 4090's 1008 GB/s, which directly translates to fewer tokens per second on the same model. The unified memory advantage manifests when model size exceeds 24-32 GB: at that point, NVIDIA users need multi-GPU or workstation cards, while Apple Silicon users just load the model. Unified memory is also non-upgradeable, buy the memory you need upfront, because you cannot add more later. This makes used Mac purchases for LLM work tricky: finding a used Mac with exactly the right memory configuration requires patience.

M1 Through M6: Generational Bandwidth Improvements

Each Apple Silicon generation has improved memory bandwidth, directly improving LLM tokens-per-second, though several generations hide a catch: the same chip name ships on two different physical dies, with the cheapest memory configuration binned onto a slower one. M1 (2020): base 68 GB/s, Pro 200 GB/s, Max 400 GB/s, Ultra 800 GB/s with up to 128 GB, the first Apple Silicon config that could run 70B models at usable speeds. M2 (2022): 10-20% improvements across the lineup. M2 Max reaches 400 GB/s; M2 Ultra reaches 800 GB/s with up to 192 GB, the first Mac to comfortably run 70B at Q8_0. M3 (2023): the Max tier's cheapest configurations actually regress. The 48/64/128 GB M3 Max builds run a full 40-core-GPU die at 400 GB/s, same as M2 Max, but the 36 GB and 96 GB configurations use a cut-down 30-core-GPU die at only 300 GB/s, a real 25% bandwidth cut for choosing less memory. M3 Ultra reaches 819 GB/s with up to 512 GB. M4 (2024): a major jump. M4 Pro reaches 273 GB/s, up 75% from M3 Pro. M4 Max repeats M3 Max's binned-die pattern: 36 GB on a cut-down die at 410 GB/s, or 48/64/128 GB on the full die at 546 GB/s. Apple never released an "M4 Ultra", M4 Max's die lacks the UltraFusion connector needed to fuse two dies, so the Mac Studio's Ultra tier stayed on M3 Ultra through this entire generation. M5 (2025-2026): base M5 shipped October 2025 at 153 GB/s, up 27.5% from M4's 120 GB/s, the largest single-generation jump in the base tier's history. M5 Pro and M5 Max followed in March 2026 for the MacBook Pro: M5 Pro reaches 307 GB/s and M5 Max repeats the binned-die split once more, 460 GB/s at 36 GB or 614 GB/s at 48/64/128 GB. M6 (2026): Apple's first 2 nm chip, replacing M4 in a new Mac mini at 170 GB/s, up 42% from M4's 120 GB/s (a more modest 11% step from base M5's 153 GB/s, the closer generational comparison since M6 is the direct successor to the M5 numbering, not to M4). It's also the first Apple chip with three distinct CPU core types (2 super cores, 4 performance cores, 6 efficiency cores) instead of two. The chart below tracks the Max tier's bandwidth across every generation, the clearest single line for what most LLM-focused buyers actually purchase.

What's New: M6 and M5 Ultra (August 2026)

On August 25, 2026, Apple refreshed both ends of the Mac lineup on the same day: a new Mac mini built around the M6 chip (with M5 Pro as its memory-upgrade option), and a new Mac Studio built around M5 Max and the all-new M5 Ultra. For local LLM inference, M5 Ultra is the headline: Apple's first Ultra-tier chip with Neural Accelerators, built by fusing two M5 Max dies via UltraFusion into a genuine quad-die chip. Neural Accelerators are dedicated matrix-multiplication hardware in each GPU core, and they target a specific, compute-bound phase: prompt processing (prefill), historically Apple Silicon's biggest weakness for local LLM work, not token generation. Apple's own MLX research on the M5 generation reports roughly 3.6-4x faster prefill than M4, but only a modest 20-27% faster token generation, since decode stays memory-bandwidth-bound regardless of how much matmul hardware is available. Getting the prefill gain also requires MLX specifically and macOS 26.2 or later; llama.cpp's Metal backend, which is what Ollama still falls back to below 32 GB of unified memory, hadn't added Neural Accelerator support as of mid-2026. It ships with up to 512 GB of unified memory (the same ceiling as M3 Ultra) at 1.2 TB/s of bandwidth, 50% more than M3 Ultra's 819 GB/s. At that top configuration, a single Mac Studio can hold essentially any open-weight model released to date, including 400B+ parameter MoE (mixture-of-experts, a model built from many specialized sub-networks instead of one dense block) models like MiniMax M3 at a near-lossless quantization. The base 96 GB configuration ships on a cut-down 64-core-GPU die, still at the full 1.2 TB/s bandwidth since Apple doesn't cut bandwidth for the smaller die this generation, and starts at $5,499; the 512 GB configuration costs more and ships separately, in late October 2026. On the budget end, the new M6 Mac mini starts at $899 with 16 GB of unified memory (configurable to 32 GB) at 170 GB/s, replacing base M4 as Apple's cheapest current Mac. It isn't a large-model machine, but it's a genuine, current entry point for 7-9B models, and the same Mac mini's M5 Pro option (up to 64 GB, 307 GB/s, from $1,699) reaches the same 27-32B sweet spot the standalone M5 Pro MacBook Pro already covers.

Which Chip for Which Model: A Practical Guide

Unified memory capacity, more than any other spec, decides which models a given Mac can actually run; bandwidth only decides how fast. The ladder below shows what each memory tier realistically unlocks, from an 8 GB Mac (not a real recommendation today) up to the new 512 GB M5 Ultra.

1. 8 GBNot a practical recommendation: this site's own OS-memory reservation leaves almost no headroom for a model. Fine for brief experimentation only.
2. 16 GB7-8B comfortably at Q5_K_M-Q6_K; 14B models only reach Q2_K here, workable but noticeably compressed.
3. 24 GB27-32B at Q3_K_M (e.g. Qwen 3.6 27B), the entry point for genuinely capable dense models.
4. 48 GB70B dense models at Q2_K, a heavily compressed but workable 70B, or 27-32B at full Q8_0 near-lossless quality. The most cost-effective config for a 70B-class model.
5. 128 GB70B at Q8_0 near-lossless, or 100-130B MoE models like Qwen 3.5 122B-A10B at Q6_K, well above the usual Q4 baseline.
6. 512 GB (M5 Ultra)Virtually any open-weight model released to date, including 400B+ parameter MoE models at a high-quality quantization.

MLX vs Ollama on Mac: Which to Use in 2026

For most of Apple Silicon's LLM history, running a model through Ollama meant running it on llama.cpp's Metal backend: solid and zero-configuration, but slower than Apple's own MLX framework, especially at prompt processing. That changed in 2026. Ollama previewed MLX as an alternative engine on March 30, 2026, and made it the default engine on Macs with more than 32 GB of unified memory as of Ollama 0.30 (May 13, 2026); Macs at or under 32 GB still get the llama.cpp Metal backend automatically. Per Ollama's own published benchmark (Qwen3.5-35B-A3B, an MoE model, on M5-generation chips), the new MLX engine decoded 93% faster and processed prompts 57% faster than the prior llama.cpp-based release. That specific comparison bundles two changes at once, the engine (llama.cpp to MLX) and the quantization format (Q4_K_M to NVFP4), so it overstates what switching engines alone buys at a fixed quantization; Ollama has not published an engine-only, same-quant version of that test. Apple's own MLX research on the M5 generation, holding quantization fixed, found a much narrower decode gain, roughly 20-27% faster token generation from the engine and its Neural Accelerators alone, against a 3.6-4x gain on prefill; that split is a better estimate of what MLX itself buys at matched quantization than Ollama's headline 93% decode figure, most of which is likely the quantization-format change. Independent benchmarks published around the same time, isolating the engine change at a matched quantization, found a similar pattern on other model sizes, summarized below: MLX's advantage is largest on MoE models, where its tighter integration with Apple's unified memory architecture pays off most. Practically, this means the old advice, use Ollama for convenience and switch to MLX only if you need the speed, is outdated for anyone on a 32 GB+ Mac: current Ollama versions already run on MLX by default. MLX is still worth using directly (via pip install mlx-lm) if you want Python-level control, LoRA fine-tuning (not available through Ollama), or you're on a Mac at or under 32 GB, where Ollama still defaults to llama.cpp.

# Ollama on Mac (simplest; uses MLX automatically above 32GB unified memory)
brew install ollama
ollama serve
ollama run llama3.3:70b

# MLX directly on Mac (Python control, needed for fine-tuning)
pip install mlx-lm
mlx_lm.generate \
  --model mlx-community/Llama-3.3-70B-Instruct-4bit \
  --prompt "Explain quantum computing"

# MLX fine-tuning (unique Mac capability, not available through Ollama)
pip install mlx-lm
# See mlx-community fine-tuning examples on GitHub

Mac vs PC for LLMs: The Honest Comparison

The Mac vs PC decision for local LLMs comes down to one question: do you need to run models larger than 32 GB? If yes, and you want to do it on a single device with minimal complexity, get a Mac with 48+ GB unified memory. If no, and you primarily run 7-32B models, a PC with an NVIDIA GPU is faster and often cheaper. Mac advantages: single-device large model capability (70B at Q4 on a MacBook Pro is genuinely unique), near-silent operation (even under sustained inference load), dramatically lower power consumption (an M4 Max MacBook Pro draws under 100W running inference, dual RTX 3090s draw 700W), portable (run 70B models on a laptop, impossible on any PC laptop), and zero configuration (Metal acceleration works out of the box, no driver installation, no CUDA toolkit). PC advantages: faster token generation for same-size models (RTX 4090 at 1008 GB/s vs M4 Max at 546 GB/s, roughly 2x faster), upgradeability (add more RAM, swap GPU, add a second GPU), gaming and other GPU workloads, CUDA ecosystem breadth (ExLlamaV2, TensorRT-LLM, and other CUDA-exclusive tools), and generally lower cost for equivalent compute. For a dedicated inference appliance: Mac Mini M4 Pro 48GB at $2399 is compelling. For a versatile workstation: PC with RTX 3090/4090/5090. For portable large-model inference: MacBook Pro M4 Max with 64-128 GB. There is no universal winner, the right choice depends on which models you run and where you run them.

Which Mac to Buy for LLM Inference

The concrete picks below cover every realistic budget, from the new entry-level Mac mini to the new Mac Studio flagship. The one rule that cuts across all of them: always maximize memory over other specs. A base M4 Pro with 48 GB is far more useful for LLMs than a maxed-out M4 Pro with 24 GB. Memory determines which models you can run at all; everything else, GPU core count included, only affects how fast they generate.

What Tokens Per Second to Expect on Apple Silicon

Token generation speed on Apple Silicon is bandwidth-bound. Approximate tok/s for common configurations, on dense models with a standard llama.cpp/Metal-class engine: 7B Q4_K_M (~5 GB): M1 8-15 tok/s, M2 Pro 15-25 tok/s, M4 Max 30-50 tok/s. 14B Q4_K_M (~9 GB): M1 Pro 8-15 tok/s, M4 Pro 18-28 tok/s, M4 Max 25-40 tok/s. 27-32B Q4_K_M (~17-20 GB): M2 Pro 5-8 tok/s, M3 Max 10-18 tok/s, M4 Max 15-25 tok/s. 70B Q4_K_M (~40 GB): M1 Ultra 12-18 tok/s, M3 Max 6-8 tok/s, M4 Max 9-10 tok/s (M1 Ultra's 800 GB/s outruns both despite being the oldest chip here, since bandwidth, not generation, is what decode speed tracks). These are dense-model figures; MoE models decode meaningfully faster per active parameter, and on a 32 GB+ Mac running current Ollama (MLX by default, see above), expect decode speeds well above this table for MoE models specifically, sometimes 2-3x, per the MLX vs Ollama benchmarks earlier in this guide. These numbers assume the model fits entirely in unified memory with all layers on GPU (which Ollama and MLX both handle automatically). Prompt processing (prefill) works differently: it's compute-bound rather than bandwidth-bound, and it's historically Apple Silicon's biggest weak point for local LLM work, not raw decode speed. A short prompt is fast on any of these chips, a 2000-token prompt processes in roughly 1-3 seconds on M4 Max, but prefill time scales up sharply with context length in a way decode doesn't: pre-M5 Ultra-tier chips working through 64K-128K token contexts have been widely reported taking several minutes just to reach the first output token, the most common real-world complaint about running large models locally on a Mac. The M5 generation's Neural Accelerators target exactly this (see above), but only for MLX and only on macOS 26.2+. If your workload leans on long documents, large codebases, or long chat histories rather than short prompts, budget for prefill time separately from the tok/s figures above, it's often the larger share of total wait time on Apple Silicon. The tok/s numbers matter most for interactive use: 10 tok/s is readable, 20 tok/s feels responsive, 30+ tok/s feels fast. Long context reduces generation speed slightly on Apple Silicon due to KV cache management, but less dramatically than on discrete GPUs because the unified memory can accommodate very large KV caches without overflow. For maximum context on Mac, ensure you have enough total memory for model + KV cache. A 70B model at Q4 (~40 GB) with 32K context needs roughly 10-12 GB for KV cache at FP16, plan for 52-55 GB total.