Best Local LLMs for Mac in 2026 — M1 through M5 Tested

📚 More on this topic: Qwen 3.6 Setup Guide · Qwen 3.5 Setup Guide · DeepSeek V4 Flash vs Pro · Running LLMs on Mac M-Series · Qwen 3.5 on Mac: MLX vs Ollama · llama.cpp vs Ollama vs vLLM · VRAM Requirements · Run 31B Models on a Laptop

Every Mac with Apple Silicon can run local LLMs. The question isn’t whether — it’s which model, and whether it’ll be fast enough to actually use. A model that “fits” in memory but generates 3 tokens per second isn’t useful. A smaller model at 40 tok/s is.

This guide gives you specific model recommendations for every Mac tier, with performance numbers drawn from community reports. We run our own benches on a 3090, not Apple silicon, so where a number comes from Mac hardware, it’s cited. No “it depends” here — concrete picks you can install right now. Updated July 2026 with the collapse of Apple’s high-memory Macs (the Studio now caps at 96GB, and the discontinued Mac Pro took the 192GB tier with it), Ollama’s matured native MLX runner, and the reframe that matters most on Mac: bandwidth, not RAM, sets your tokens per second.

For setup instructions and general architecture details, see our complete M-series guide. This article focuses on which models to run.

What changed since the last refresh

The short version: the MoE revolution finally landed on Mac.

• Qwen 3.6-35B-A3B (released April 16, 2026) is a 35B MoE with only 3B active per token. That math changes everything on Mac — the model file is ~20GB at Q4, but token-generation speed feels like a 3B model. Native context 262K, extendable to ~1M via YaRN, Apache 2.0.
• Qwen 3.6-27B dense (released April 22, 2026) is the new flagship dense coding model. 77.2 on SWE-bench Verified. Apache 2.0. Simon Willison ran the 16.8GB Unsloth Q4_K_M GGUF and clocked 25.57 tok/s — flagship-class output on a single Mac.
• Gemma 4 26B-A4B landed with 4B active params and 256K context under Google’s Gemma license. Good alternative MoE if you want something other than Qwen.
• DeepSeek V4-Flash (284B total, 13B active, 1M context, MIT — see our V4 Flash vs Pro guide) is Mac-runnable at aggressive quant on 128GB+ configs, and the Mac benchmarks that were missing through mid-2026 have now landed — community MLX quants (inferencerlabs, Deviad) run it at ~27–34 tok/s at IQ2 on a 128GB Mac, ~26–37 at 2–4-bit on a 192GB+ Ultra.
• The high-memory Macs didn’t get delayed — they got deleted. The global DRAM shortage, driven substantially by local-AI demand (reporting names the early-2026 OpenClaw boom specifically), gutted Apple’s lineup. The 512GB M3 Ultra Studio died in March 2026, 256GB was gone by May, and the Studio now tops out at 96GB. The 192GB Mac Pro was discontinued outright on March 26. The 128GB M5 Max MacBook Pro is now the highest-memory Mac you can buy, of any kind. If a guide told you to buy a 256GB Studio, that machine no longer exists — so the tier tables below split what you can still buy from what you might already own.

Qwen 3.5 is still fine. If you’re on Ollama and hit the qwen35moe mmproj bug or haven’t upgraded your tooling for 3.6 yet, Qwen 3.5-9B and 3.5-32B remain solid picks. But the defaults have moved.

And no, 3.6 isn’t already outdated. The newer Qwen 3.7-Max and 3.8-Max (previewed July 19, 2026) are API-only — no open weights — and Moonshot’s Kimi K3 ships around 1.4TB at 4-bit, too big for any Mac, including the discontinued 512GB Studio. Qwen 3.6 is still the open-weight default you can actually download and run.

Why Mac Is Different

Unified Memory Changes the Math

On a PC, your GPU has its own dedicated VRAM (typically 8-24GB). Models that don’t fit in VRAM either won’t run or crawl at 2-3 tok/s via offloading.

On Mac, there’s no separate GPU memory. Your entire RAM pool — 8GB up to the 128GB ceiling — is shared between CPU and GPU. A Mac Mini with 48GB can load a 32B model that would need a ~$1,000+ used RTX 3090 on PC (the 2026 memory crunch pushed used-3090 prices up from the $700s a year ago). A 96GB Mac Studio, or a 128GB M5 Max MacBook Pro, runs 70B models that require $3,000+ in dual GPUs on PC.

The tradeoff: Mac’s memory bandwidth is lower than a discrete GPU’s. An RTX 3090 pushes 936 GB/s. The M4 Pro pushes 273 GB/s. Token generation speed is directly proportional to memory bandwidth, so Mac is 30-60% slower per token for models that fit in a GPU’s VRAM. But for models that don’t fit — Mac wins by running them at all.

MoE changes this equation. Qwen 3.6-35B-A3B only activates 3B params per token. That means token speed is closer to a 3B dense model’s, even though the file is 35B’s worth of disk and memory. This is why the new Qwen 3.6 MoE is such a big deal on Mac specifically.

Memory Bandwidth Matters More Than Chip Generation

This is the counterintuitive part, and it’s the single most useful thing to understand before you spend a dollar. RAM tells you if a model loads. Bandwidth tells you how fast it runs. An M3 Max (400 GB/s) generates tokens faster than an M4 Pro (273 GB/s) despite being a generation older, because for LLM inference bandwidth is the bottleneck, not compute. Every Mac buying guide leads with memory size and buries bandwidth. That’s backwards.

Here’s what that looks like across the machines you’d actually shop today — and it’s checkable arithmetic, not a benchmark we ran:

Read that top to bottom and the trap jumps out: the 96GB M3 Ultra moves tokens faster than the 128GB M5 Max, even though it holds less. More RAM, slower generation. If your model fits in 96GB, the Ultra is the faster machine.

The M5 Max trap specifically: “M5 Max” is not one number. The full 40-core-GPU chip does 614 GB/s. The binned 32-core version in the base 14-inch ($3,599) does 460 GB/s — that’s 56% of the M3 Ultra’s bandwidth, not the 75% you’d get from the 614 chip. Someone reaches for “M5 Max, 128GB” expecting the fast one and pays laptop money for the slow one. Check the GPU core count before you buy, not just the chip name.

For the chip you already own, here’s the generational picture:

Before buying: check the bandwidth of your specific chip, not just the generation. A Mac Mini M4 Pro 48GB is slower per token than a Mac Studio M4 Max 64GB, even on the same model.

Best Models by Mac Tier

8GB Macs (M1 / M2 / M3 / M4 base)

macOS needs 2-3GB for itself. You have about 5-6GB for a model. This limits you to 3B-4B models comfortably or 7B-8B models with aggressive quantization and short context.

The pick: Gemma 4 E2B if your tool supports it, Qwen 3.5 4B as the safe fallback. Both fit easily, leave room for context, and stay fast.

Skip: Any 7B+ model at Q4 or higher. It’ll technically load but you’ll have 1-2GB for context and system, which means frequent crashes and 4K token limits.

Honest take: 8GB Macs are getting uncomfortable in 2026. 2026’s model generation — Qwen 3.6, Gemma 4 26B-A4B, DeepSeek V4 — is explicitly built for MoE configs with more memory headroom. If you’re serious about local AI, the memory upgrade is worth it. An M4 or M5 MacBook Air starts at 16GB now — that’s the minimum to buy going forward.

16GB Macs (M1 Pro 16GB / M2 16GB / M3 16GB / M4 16GB / M5 16GB)

The 7B-8B tier. You have ~12-13GB available for the model and context.

The pick: Qwen 3.5 9B (Q4_K_M). This has held its slot since the last update — fits in ~6.6GB via Ollama, beats models 3x its size on reasoning, /think mode when you need chain-of-thought. See our 9B setup guide.

Worth testing: Gemma 4 E4B. Google’s smaller Gemma 4 variant has picked up steam on the Arena leaderboard. Worth a try if you want an alternative to Qwen.

Honest take: 16GB Macs are now the floor for useful local AI. You can run capable 8B-9B models, but the 32B+ tier with Qwen 3.6-35B-A3B is where 2026 gets interesting — and that requires 32GB+. If you’re buying new, step up.

24GB Macs (M2 Pro 24GB / M4 Pro 24GB / M4 16GB with swap)

The 14B tier and the low edge of the Qwen 3.6-27B zone.

The pick: Qwen 3.6-27B (Unsloth Q4_K_M GGUF). Per Simon Willison’s April 22 post, this runs at 25.57 tok/s on his machine with the 16.8GB GGUF — flagship-class coding output on a single Mac. On a 24GB Mac you’re at the knife’s edge: model is 16.8GB, leaving ~5GB for macOS and context. Expect shorter context limits and occasional memory pressure. Good tradeoff if coding is the workload.

Safer pick: Qwen 3 14B. If you don’t want to live on the edge, 14B at Q4 leaves comfortable context headroom and still delivers strong general output.

Don’t bother with: Qwen 3 32B at Q3. It technically fits but quality at Q3 is degraded enough that a 14B at Q4 usually wins. Wait for more memory.

32-48GB Macs (M3 Pro 36GB / M4 Pro 48GB / M2-M4 Max 32-48GB)

This is where Qwen 3.6-35B-A3B MoE wins outright. The model fits, the active-param footprint is small, and you get room for context.

The pick: Qwen 3.6-35B-A3B (Q4_K_M via Ollama, MLX 4-bit via MLX-LM). This is the model that makes Mac local AI worthwhile in 2026. MoE design means 3B active params per token, so token speed stays high even as the 20GB file sits in memory. Native 262K context. Strong on code (73.4 SWE-bench Verified), strong on reasoning (92.7 AIME26), strong on general tasks (85.2 MMLU-Pro).

MLX variants are live at mlx-community/Qwen3.6-35B-A3B-4bit and unsloth/Qwen3.6-35B-A3B-UD-MLX-4bit.

For coding specifically: Qwen 3.6-27B dense. Higher single-token quality than the A3B MoE on code. Slower per token (dense model hits full 27B per step), but the output is worth it. Simon Willison’s verdict on the Q4 GGUF: “an outstanding result for a 16.8GB local model.”

Alternative MoE: Gemma 4 26B-A4B. 4B active, 256K context, different training recipe. Worth having on disk for variety.

The Mac Mini M4 Pro 48GB remains the best-value setup in this tier — though “best value” now costs more: Apple raised the M4 Pro mini’s starting price by $200 in June 2026, and a 48GB config lands around $1,999. Silent, low power, runs Qwen 3.6-35B-A3B all day with room for 32K+ context. It’s still the machine to beat under two grand.

48-64GB Macs (M3 Max 48-64GB / M4 Max 64GB / M5 Max 48-64GB)

70B models become practical. Qwen 3.6-35B-A3B runs with generous context. You can keep multiple models resident.

The pick: Qwen 3.6-35B-A3B at Q8. At this memory tier you can afford the higher quant. Quality improvement over Q4 is meaningful on edge cases, and the MoE architecture keeps it fast.

For coding: Qwen 3.6-27B at Q6 or Q8. Pairs the best dense coding model with high-quality quantization.

If you want a 70B: Llama 3.3 70B is still the standard at Q4. 8-15 tok/s on M4 Max — slower than reading speed but usable for interactive work.

96-128GB Macs (M3 Ultra 96GB, M5 Max MacBook Pro 128GB)

The “no compromises” tier for consumer hardware. Frontier-class models become practical. Buying note: this tier changed shape in 2026. The 128GB M4 Max Studio is gone (that chip caps at 64GB now), so a desktop at this level means the 96GB M3 Ultra Studio at $5,299. For 128GB you’re now looking at the M5 Max MacBook Pro — the only Mac that still reaches it.

The pick: Qwen 3.6-35B-A3B at bf16 if you want to see what full-precision Qwen 3.6 looks like on a consumer Mac. This is as close to the Hugging Face Inference API experience as you’ll get locally.

DeepSeek V4-Flash — now benchmarked, no longer just aspirational. The 284B/13B-active MoE has real Mac numbers as of late July: community MLX quants (inferencerlabs, Deviad) run it at ~27–34 tok/s on a 128GB Mac — but that speed rides on IQ2 (2-bit), which trades real quality for the fit. The 4-bit weights are ~140GB and don’t fit 128GB; for 4-bit you need a 192GB+ Ultra, where it lands ~26–37 tok/s. So on a 128GB Mac it’s genuinely usable now, just at aggressive quant. See our V4 Flash vs Pro guide for sourcing.

Llama 4 Scout is the practical MoE pick at this tier. 109B total / 17B active fits in ~58 GB at Q4, leaving room for context on 96GB+ Macs. M5 Max users have reported ~32 tok/s via MLX. MoE caveat: Scout’s expert-routing degrades more on aggressive quant than dense models — the router operates on weight distributions, and Q4 loses enough precision to pick wrong experts more often. If memory allows, run Q5 instead of Q4. At 96GB you have the headroom for it; at 128GB it’s a no-brainer.

Skip DeepSeek V4-Pro. 1.6T total / 49B active. Even at aggressive quant it’s ~800GB of weights. Not a Mac story, period.

192GB+ Macs (owner-only in 2026 — you can no longer buy in)

The rarefied tier, and the one the memory shortage erased. Here’s the sequence: Apple killed the 512GB M3 Ultra around March 6, 2026, citing the DRAM shortage; 256GB was gone by May; and the Studio now caps at 96GB. The 128GB option went at some point in between — Apple never announced when. The 192GB Mac Pro that could have anchored this tier was discontinued entirely on March 26.

So if you already own a 192GB M2 Ultra or a 256GB M3 Ultra you bought before the cuts, everything below still applies to your machine. But if you’re shopping today, this tier is closed. The ceiling you can actually buy is a 96GB Studio or a 128GB M5 Max MacBook Pro. Any guide still telling you to “get the 256GB Studio for frontier models” is pointing at a machine that isn’t for sale.

The pick: DeepSeek V4-Flash is the most interesting model to try at this tier. 1M native context window, MIT licensed, and the 13B active-param footprint keeps token speed reasonable once weights are loaded. MLX quants have shipped (inferencerlabs, Deviad) — check the mlx-community org and individual uploaders.

MLX quants of the V4 architecture have shipped (inferencerlabs, Deviad, and others) — Ivan Fioravanti’s MLX context benchmarks on an M3 Ultra put it around 26–35 tok/s at 2–4-bit. If you have a 192GB Ultra and you’re not experimenting with 200B+ parameter models, what are you even doing with it.

MLX vs Ollama vs LM Studio in 2026

All three work on Apple Silicon. The difference is speed, ease, and which models are supported on day one.

Numbers above are approximate community reports on Qwen 3.6-35B-A3B Q4 on M4 Max configs. Individual mileage varies substantially by RAM size, concurrent load, and quant variant.

One thing that changed in 2026: Ollama switched to a native MLX runner on Apple Silicon (introduced with 0.19 on March 31, 2026), so on a modern Mac it’s no longer just a llama.cpp wrapper — it taps the same Apple framework MLX-LM does. It’s kept improving since: Ollama’s MLX engine now uses the M5’s GPU Neural Accelerators for faster time-to-first-token, fuses operations into single Metal kernels for another ~20% on some workloads, and supports NVFP4 4-bit quants. That narrows the old speed gap; the remaining difference is mostly convenience versus squeezing out the last few tok/s.

When to Use MLX

MLX is Apple’s native machine learning framework, built for unified memory. On supported models it’s typically 10-30% faster than llama.cpp. The gap used to be larger — llama.cpp has closed a lot of ground on Mac since 2025, and for some recent models at Q4 GGUF the gap is near-zero.

Use MLX when:

• You want maximum tok/s and you’re comfortable with Python
• You’re building apps that need fast local inference
• You’re on Qwen 3.6 or Gemma 4 specifically — MLX Community has full 3.6 coverage now

pip install mlx-lm
mlx_lm.generate --model mlx-community/Qwen3.6-35B-A3B-4bit --prompt "Hello"

The mlx-community HF org maintains MLX-converted variants of nearly every mainstream model within days of release. Unsloth also ships MLX builds now — see unsloth/Qwen3.6-35B-A3B-UD-MLX-4bit. For a head-to-head speed comparison on the Qwen 3.5 family specifically, see our MLX vs Ollama benchmark on Apple Silicon.

When to Use Ollama

Ollama wraps a dead-simple model manager over its inference engines — MLX on Apple Silicon (32GB+), llama.cpp elsewhere. One command to install, one command to run.

Use Ollama when:

• You want the fastest setup possible
• You need an API server for other apps (Open WebUI, Continue, etc.)
• You’re new to local LLMs
• You want to switch between models quickly

ollama run qwen3.6:35b-a3b-q4_K_M

Known Ollama gotcha: the qwen35moe mmproj bug, filed March 9, 2026 and closed as a duplicate of #14575, affects Qwen 3.5 MoE variants when used with separate vision projector files. Ollama errors with “unknown model architecture: ‘qwen35moe’”. Workaround: use llama.cpp directly with --mmproj, or switch to Qwen 3.6 for vision work (different arch, not affected by this specific bug). Monitor the tracking issue before committing to a vision pipeline on Qwen 3.5 MoE via Ollama.

When to Use LM Studio

LM Studio gives you a ChatGPT-like interface with model browsing, parameter controls, and conversation management. Recent versions use MLX as the backend on Mac when available, closing the speed gap with MLX-LM — the 0.4.13 release (May 2026) moved to MLX engine v1.8.1 and added parallel predictions for vision models like Qwen 3.6 and Gemma 4. Current is 0.4.20 (July 22, 2026), which also ships a separate agent app, LM Studio Bionic.

This is the tool Simon Willison used to run Qwen 3.6-35B-A3B on his MacBook Pro M5 with the UD-Q4_K_S 20.9GB GGUF. His verdict after the pelican-on-a-bicycle test: “I’m giving this one to Qwen 3.6. Opus managed to mess up the bicycle frame!” — a light result, but one he didn’t expect.

Use LM Studio when:

• You prefer a GUI over terminal
• You want to browse and compare models visually
• You need fine control over temperature, top-p, and sampling
• You want MLX backend without writing Python

DeepSeek V4 on Mac: the honest take

DeepSeek V4 dropped the evening of April 23, 2026 — two variants, MIT license, 1M context. See our V4 Flash vs Pro guide for the full breakdown. For Mac specifically:

V4-Pro (1.6T / 49B active). Not a Mac story. Even on a maxed Ultra, the weights don’t fit without disk offload, and disk offload at 1M context is not going to be fun. Server hardware only.

V4-Flash (284B / 13B active). Mac-runnable now, with real numbers. Community MLX quants (inferencerlabs, Deviad) run it at ~27–34 tok/s on a 128GB Mac at IQ2 — 2-bit, so a genuine quality trade for the fit. The 4-bit weights are ~140GB and need a 192GB+ Ultra, where it lands ~26–37 tok/s. So 128GB is usable at aggressive quant; 192GB+ gets you 4-bit comfortably. Check the mlx-community org and individual uploaders for the current quants.

What to actually do: if you’re on 96GB+ and curious, monitor the mlx-community HF org for V4-Flash MLX uploads and check the V4 Flash vs Pro guide for the latest sourcing. If a working quant isn’t up yet, don’t burn time converting the 284B checkpoint yourself.

M5 Max and the Mac Studio Situation

M5 Max MacBook Pro shipped in March 2026. Bandwidth is up substantially from M4 Max — the M5 Max / M5 Pro guide covers the numbers. For local LLM work specifically, the bandwidth bump matters more than the compute bump.

The complication is bigger than a delay. The same DRAM shortage that pushed the M5 Studio refresh back (Bloomberg’s Mark Gurman reported an October 2026 target at the earliest) also deleted the high-memory desktop configs that made the Studio interesting for big models. The Studio caps at 96GB now. The 192GB Mac Pro is gone. So the desktop question isn’t “wait for M5 or buy M4 now” — it’s “96GB is the desktop ceiling until further notice, whatever chip is inside.”

What this means practically:

• Need a desktop for local AI now? The buyable options are a 96GB M3 Ultra Studio ($5,299) or a smaller M4 Max Studio. There is no higher-memory desktop to wait for in the near term.
• Need more than 96GB? Your only path is the 128GB M5 Max MacBook Pro. The laptop, not the desktop, holds the memory crown in 2026.
• Delivery: don’t plan around a fixed wait. It’s config-dependent, and the multi-month backorders you may have read about vanished — not because supply recovered, but because the configs people were waiting on (256GB, 512GB) no longer exist. Check Apple’s live shipping estimate for the exact build before you commit.

Don’t treat rumored Ultra variants as fact — Gurman’s talk of a much-higher-memory M5 Ultra is a rumor, and the M3 Ultra Studio itself was a surprise release. Plan around what’s shipping, not what’s leaked.

Models That Technically Fit But Actually Crawl

This is the trap. A model can load into memory and still be useless.

The rule of thumb: the model file should be no more than 60-70% of your total memory. That leaves room for macOS, the KV cache (context), and framework overhead. A 20GB model on a 48GB Mac is comfortable. A 20GB model on a 24GB Mac is a knife’s edge.

If you’re right at the limit, drop to a lower quantization or a smaller model. A snappy 14B model is more useful than a sluggish 32B.

Want to check the math for your specific Mac? Use the Local AI Planning Tool — pick your Mac’s unified memory, see every model that fits at your quant, with per-model VRAM breakdowns including KV cache at your target context length.

The Best Mac for Local AI in 2026

These are 2026 prices, after the June memory-driven hikes. Everything moved up, and the top of the table lost its high-memory rungs entirely.

Two things that table makes obvious. First, the Mac Mini M4 Pro 48GB — now around $1,999 — is still the sweet spot. It runs Qwen 3.6-35B-A3B comfortably, sits silently on your desk, draws about 30W under AI load, and costs less per year in electricity than a single month of ChatGPT Plus. Second, the memory crown and the speed crown split in 2026: the 128GB laptop holds the most RAM, but the 96GB M3 Ultra Studio moves tokens faster. Pick by which one your model actually needs. (Apple doubled the 128GB RAM-upgrade price this year, which is how a maxed 16-inch now reaches five figures.)

If you’re shopping a laptop and want the memory ceiling, it’s the M5 Max MacBook Pro at 128GB — just buy the 40-core-GPU chip, not the 460 GB/s binned one.

The Bottom Line

The model matters more than the tool. Pick the right model for your memory tier, use whichever app you’re comfortable with, and don’t try to squeeze a model that’s too big. A fast small model beats a slow big one every time.

Quick decision tree for 2026:

• 8GB: Gemma 4 E2B or Qwen 3.5 4B. Accept the limitations. The platform is outgrowing you.
• 16GB: Qwen 3.5 9B. Still useful, but the interesting 2026 models need more memory.
• 24GB: Qwen 3.6-27B Q4 if you can tolerate tight context; Qwen 3 14B if you want comfort.
• 32-48GB: Qwen 3.6-35B-A3B is the 2026 default. Fast MoE, strong all around, fits comfortably with 32K+ context.
• 48GB+: Qwen 3.6-27B dense at Q6/Q8 is the best local coding model. Test Simon Willison’s setup.
• 64-96GB: Qwen 3.6-35B-A3B at Q8 plus a 70B resident. Gemma 4 26B-A4B for variety. Llama 4 Scout becomes plausible at 96GB.
• 128GB (M5 Max MacBook Pro): DeepSeek V4-Flash now runs at ~27–34 tok/s (IQ2 — a 2-bit quality trade). Usable if you accept aggressive quant; 4-bit needs a 192GB+ Ultra. This is now the memory ceiling for anything you can buy.
• 192-256GB Ultra (owner-only): DeepSeek V4-Flash comfortably — if you already own one. You can’t buy into this tier anymore; the shortage deleted it.

Install Ollama (curl -fsSL https://ollama.com/install.sh | sh) or MLX-LM (pip install mlx-lm), pull your model, and start chatting. Your Mac is already a capable AI workstation — you just need to pick the right model for it.

The defaults moved in April 2026. If you haven’t pulled Qwen 3.6 yet, that’s the weekend project.

Last updated July 23, 2026. July 23 refresh: DeepSeek V4-Flash now has real Mac benchmarks (~27–34 tok/s at IQ2 on a 128GB Mac, ~26–37 at 2–4-bit on a 192GB+ Ultra) and shipped MLX quants (inferencerlabs, Deviad), so its tier picks moved from “aspirational” to “usable at aggressive quant”; bumped LM Studio to 0.4.20 (plus the new Bionic agent app); and confirmed Qwen 3.7/3.8-Max are API-only, so 3.6 stays the open-weight default. Earlier (July 17): rebuilt around the 2026 memory collapse — the Mac Studio now caps at 96GB (256GB and 512GB discontinued), the 192GB Mac Pro is gone, the 128GB M5 Max MacBook Pro is the new memory ceiling — with the bandwidth-over-RAM framing and the M5 Max 460-vs-614 GB/s trap.