Oct 21, 2026

Ollama v0.40 Makes MLX the Default on Every Apple Silicon Mac: What Changed, the Disk Gotcha, and Which Mac to Buy (October 2026)

Q: How do I know MLX is actually being used?

`ollama ps` after loading a model — the RUNNER column reads `mlx` or `llamacpp`. Don't trust vibes; a surprising number of "MLX made nothing faster" complaints turn out to be GGUF tags that never left llama.cpp.

By RunAIHome Team · 14 min read

ollamamlxapple-siliconmaclocal-llmhardware

TL;DR: Ollama v0.40.0 (September 25, 2026) flipped the MLX selection rule on Apple Silicon: models whose architecture the MLX runtime supports — Gemma 4, Qwen3.6, Qwen3.5, the decision-model family, embeddinggemma-2, with more being enabled — now run on MLX automatically, and everything else falls back to llama.cpp. The old 32GB-Mac gate from the v0.30 era is gone from the release notes; the gate is now your model, not your memory. Two things to act on: check ollama ps to see which runner you’re actually on, and know that v0.40.2’s background model upgrades keep full backups on disk until you delete them. For buyers, MLX-by-default doesn’t change the physics — decode speed still tracks memory bandwidth — so the $1,500–$3,000 question comes down to the Mac Mini M5 Pro’s 307 GB/s versus the base Mac Studio M5 Max’s 460.8 GB/s.

	Mac Mini M5 Pro 64GB	Mac Studio M5 Max 36GB	Used RTX 3090 24GB
Memory / bandwidth	64GB @ 307 GB/s	36GB @ 460.8 GB/s	24GB @ 936 GB/s
Price (Oct 2026)	$2,699–$2,899	$2,499	$1,456–$1,558 (card only)
The catch	Half the Studio’s bandwidth	36GB caps you below dense 70B	Loud, 285–350W, no macOS

Honest take: v0.40 is the first Ollama release where a 16GB MacBook Air gets the same MLX path as a Mac Studio, and that’s genuinely new — but it’s a software convenience, not a hardware upgrade. If you’re buying a Mac for local AI in this budget, get the $2,499 Mac Studio M5 Max when your models are 27–35B-class and speed matters, and the Mac Mini M5 Pro at 64GB when you specifically want dense-70B capacity in a quiet box. If macOS isn’t the point, a used RTX 3090 is still 2–3× faster than either on everything it can hold.

Ollama has been circling MLX all year: the v0.19 preview in March brought Apple’s framework in behind a flag with a 32GB floor, and v0.30 in May stabilized it as the default fast path on capable Macs. The v0.40 line, shipping September 25 through October 18, finishes the job — and quietly changes who benefits, how your models are stored on disk, and what to check before you blame your Mac for slow tokens.

What v0.40 actually changed

The selection rule. Per the official release notes, v0.40.0 says model architectures supported by the MLX runtime “will automatically run on MLX” on Apple Silicon, with unsupported architectures running on llama.cpp as before. The named list at launch: Gemma 4, Qwen3.6, Qwen3.5, the decision models (Nimble, tev1, clef, clef-flash), and the embeddinggemma-2 embedding model — with the team stating they’ll keep enabling more.

That’s a different rule than the one we documented in June. The v0.30-era default was gated on hardware — our testing notes and Ollama’s own framing put MLX on 32GB+ Macs, which excluded every 16GB and 24GB machine Apple sells. The v0.40 notes contain no memory floor at all: the gate is whether your model’s architecture has an MLX implementation. A base M6 Mac Mini or a 16GB MacBook Air pulling gemma4 now gets the MLX runner, same as a Mac Studio.

The point releases filled in the operational details:

Version	Date	What it did
v0.40.0	Sep 25	MLX by default for supported architectures; decision models + embeddinggemma-2 on MLX
v0.40.2	Oct 8	Background-upgrades previously downloaded models on first run; keeps originals as backups
v0.40.3	Oct 11	Pauses the background upgrade (it was hurting embedding/request performance)
v0.40.4	Oct 12	Improved speculative decoding in the MLX runner; decision models runnable from the CLI
v0.40.5	Oct 14	Nemotron 3.5 Super and Bolt added (Bolt: no Apple Silicon/Metal support yet)
v0.40.6	Oct 16	q2_0 tensor and file types land (the 12GB-GPU story we covered)
v0.40.7	Oct 18	Draft models for speculative decoding via llama-server; MLX and llama.cpp updates

Two of those rows are warnings in disguise, covered below.

First thing to run: find out which engine you’re on

Don’t assume. Load a model and check the runner column:

$ ollama run gemma4:12b "hello" 
$ ollama ps
NAME          ID            SIZE    PROCESSOR    RUNNER    UNTIL
gemma4:12b    a2b7c91d...   8.9 GB  100% GPU     mlx       4 minutes from now

RUNNER reads mlx or llamacpp. If you expected MLX and see llamacpp, the usual reasons, in order of likelihood:

Your model’s architecture isn’t on the MLX list yet. Llama-family models, most community GGUF imports, and anything you pulled from Hugging Face as a GGUF run on llama.cpp. That’s by design, not a bug.
You pulled an explicit GGUF quantization tag. Tags like :27b-q4_K_M or :8b-q8_0 are GGUF files and run on the llama.cpp path; the default tags are what Ollama is free to serve with MLX weights.
You’re on an old version. ollama --version should read 0.40.x; the macOS app updates itself, a Homebrew install (brew upgrade ollama) doesn’t always keep pace.

There is still no documented environment variable that forces MLX on or off globally — if a guide tells you to export some OLLAMA_MLX=1 flag, it’s describing behavior the docs don’t state. The architecture list is the switch, and Ollama controls it.

The disk gotcha: v0.40.2 keeps a backup of every upgraded model

This is the “actual problem you’ll hit” section. When v0.40.2 shipped, Ollama began upgrading models you’d downloaded under older versions in the background the first time you ran them — and it keeps the original as a backup. The release notes are explicit: upgraded models take additional disk space, and downgrading below 0.40 means re-pulling those models.

On a Mac where models live on the boot SSD and a healthy local collection runs 100–300GB, silently duplicating model blobs is how you wake up to a full 512GB drive. The symptom: your ~/.ollama/models directory is suddenly far larger than ollama list suggests, right after upgrading to 0.40.x.

The fix ships in the same release notes — a one-liner that finds and deletes the backups:

curl -s http://localhost:11434/api/tags | \
  jq -r '.models[] | select(.name | endswith(".backup")) | .name' | \
  xargs -I{} ollama rm {}

(Variants of this command circulate; the principle is the one from the notes — backups are regular entries you can ollama rm. If you might ever downgrade below 0.40, keep them: the upgraded blobs don’t load on older versions.)

Note also v0.40.3 paused the aggressive background upgrading because it degraded embedding and request performance. If your Mac felt busy at idle in the second week of October, that was why — update to 0.40.3 or later and it stops.

What MLX-by-default is actually worth in tokens per second

The honest version, assembled from measured sources rather than the “3× faster!” headlines:

The headline gains are real but model-shaped. Ollama’s own M4 Max testing during the preview showed Qwen3.5-35B-A3B going from 57.8 to 112 tok/s, prefill from 1,154 to 1,810 tok/s — a 93% decode gain on a sparse MoE model. Across the broader model zoo, measured roundups put MLX’s typical advantage at 20–40% over the llama.cpp Metal path, and it shrinks toward zero on dense 27B+ models where both engines saturate the same memory bus.
Part of the gain is Ollama fixing Ollama. The same analysis measured raw llama.cpp at 89.4 tok/s on an M4 Max where Ollama’s llama.cpp path delivered 43.5 tok/s — meaning Ollama’s old Mac stack carried heavy overhead of its own. The MLX runner shed that overhead, which is why Ollama users see bigger upgrade deltas than raw llama.cpp-vs-MLX benchmarks imply.
Long context flips the result. MLX’s decode lead comes with weaker time-to-first-token on big prompts — one October analysis found the advantage inverting around 40K tokens of context. If your workload is repo-dump agentic coding rather than chat, the MLX default helps you less than the benchmarks suggest.
v0.40.4’s speculative decoding stacks on top. Improved MLX speculative decoding plus the Gemma 4 MTP path from the v0.30 era means multi-turn chat on supported models is where a Mac feels most transformed.

So which chip gains most from the Metal→MLX switch? The answer is “whichever one runs MoE models” more than any particular silicon: sparse models with 3–4B active parameters are where MLX’s scheduling shines (that 93% M4 Max gain), while a dense 27B at Q4 barely moves on any chip. Per-chip, the gain percentage has been roughly similar up and down the line — bandwidth sets the ceiling, MLX just gets you closer to it.

The buying question: which Mac, at $1,500–$3,000

What v0.40 changes for buyers is the floor, not the ceiling: you no longer need a 32GB Mac to get the good engine. What it doesn’t change is physics — decode speed tracks memory bandwidth almost linearly (the M4 Pro at 273 GB/s measures 49.6 tok/s on a 7B Q4 in llama.cpp where the 120 GB/s M4 base measures 24.1, per the llama.cpp Apple Silicon thread). Here’s the current lineup through that lens:

Machine	Memory	Bandwidth	Price (Oct 2026)	What it means
Mac Mini M6 (base)	16GB	153 GB/s	$899	7–8B models only; ~12GB usable
Mac Mini M6 (top)	32GB	170 GB/s	$1,299	Fits 27B class at ~8–10 tok/s (est.)
Mac Mini M5 Pro	24–64GB	307 GB/s	$1,699 base; $2,699–$2,899 at 64GB	The capacity play: dense 70B Q4 fits at 64GB
Mac Studio M5 Max (32-core GPU)	36GB	460.8 GB/s	$2,499	The speed play: fastest Mac under $3,000
Mac Studio M5 Max (40-core GPU)	48–128GB	614.4 GB/s	$3,099 (48GB) and up	Above budget; where 120B MoE gets interesting

Prices and bandwidths: Apple’s August 25 launch, Macworld’s M6 Mini coverage, The Apple Wiki’s M5 Max spec table (the 36GB config rides a 384-bit bus at 460.8 GB/s; 614.4 GB/s requires the 48GB+/40-core tier), and refurb.me’s launch-price tracking for M5 Pro configs ($2,699 at 64GB with the 15-core CPU, $2,899 with the 18-core).

The decision inside the budget:

The $2,499 Mac Studio M5 Max 36GB is the speed pick. 460.8 GB/s is 1.5× the Mini M5 Pro’s bandwidth, which translates directly to 1.5× the tokens on every bandwidth-bound workload. 36GB holds the entire 27–35B class — Qwen3.8-27B at Q4 with long context, the 35B MoE daily drivers at 60+ tok/s — and the M5 generation’s per-core Neural Accelerators attack the prefill weakness that made older Macs painful on long prompts. What it can’t do: dense 70B. A 40GB Q4_K_M file doesn’t fit a 36GB machine, and no software release changes that.

The Mac Mini M5 Pro at 64GB ($2,699–$2,899) is the capacity pick. Dense 70B Q4 fits with room for context, something nothing else under $3,000 in Apple’s lineup can claim — but at 307 GB/s it decodes a 70B at roughly 7–10 tok/s (bandwidth extrapolation; reading speed, not working speed). Buy it if your use case is “the biggest model possible, quietly, in a tiny always-on box” — a Tailscale-served home AI box fits this machine’s personality. Skip the $1,699 24GB base for AI: it’s $400 more than the M6 Mini 32GB for less model capacity, and its bandwidth advantage only pays off on models both can hold.

The M6 Mini 32GB ($1,299) is the budget backstop — and v0.40 just made it slightly better, because its 24GB config class no longer misses an MLX memory gate. It’s still a ~8–10 tok/s machine on dense 27B. MoE models are its daily drivers.

And the perennial asterisk: a used RTX 3090 at $1,456–$1,558 (ResalePrices, October 20: $1,533 average across 320 listings) has 936 GB/s — double the base Studio, triple the Mini M5 Pro — and runs everything that fits in 24GB roughly 2–3× faster than any Mac in this budget. You accept 285–350W, fan noise, and a tower; see Mac vs PC for local AI for the full trade.

What to actually buy

Prices as of October 2026, all verified in the comparison above:

Your situation	The machine	Price	Where
27–35B models daily, want the fastest Mac under $3,000	Mac Studio M5 Max 36GB	$2,499	Check price
Dense 70B in a silent always-on box	Mac Mini M5 Pro 64GB	$2,699–$2,899	Check price
Cheapest Mac that fits the 27B class	Mac Mini M6 32GB	$1,299	Check price
Raw speed per dollar, macOS optional	Used RTX 3090 24GB	$1,456–$1,558	Check price
Not sure yet — test your workload on rented hardware first	Rented 3090/4090 on Vast.ai	from $0.07/hr	Check availability

Before committing, run your target model + context through our VRAM calculator — the capacity line between the 36GB Studio and the 64GB Mini is exactly where most buyers’ regret lives.

FAQ

Does Ollama v0.40 make my existing Mac faster? If you run Gemma 4, Qwen3.5/3.6, or another MLX-listed architecture and were previously on the llama.cpp path (sub-32GB Mac, or pre-0.30 install), yes — typically 20–40%, up to ~90% on MoE models. If you run Llama-family models or explicit GGUF quant tags, nothing changes; you’re on llama.cpp either way.

How do I know MLX is actually being used? ollama ps after loading a model — the RUNNER column reads mlx or llamacpp. Don’t trust vibes; a surprising number of “MLX made nothing faster” complaints turn out to be GGUF tags that never left llama.cpp.

Why did my disk fill up after updating to v0.40? The v0.40.2 background upgrade re-encodes previously downloaded models and keeps the originals as backups. Delete the .backup entries with ollama rm (command above) once you’re sure you won’t downgrade below 0.40 — downgrading requires re-pulling the upgraded models.

Is MLX always faster than llama.cpp on a Mac? No. The decode advantage is real on small-to-mid dense models and MoE; it converges to roughly nothing on dense 27B+ (both engines hit the same bandwidth wall), and time-to-first-token on very long prompts (~40K context) can favor llama.cpp. Agentic coding with huge prompts is the workload where the MLX default underdelivers.

Should I buy a Mac or a GPU for local AI in late 2026? Bandwidth decides. Under 24GB of model, a used RTX 3090’s 936 GB/s beats every sub-$3,000 Mac by 2–3×. Above 24GB, the Mac’s unified memory is the only consumer path at this budget — the 64GB Mini M5 Pro holds models no consumer GPU can. The full Mac-vs-PC breakdown has the long version, and the BYOK coding-assistant setup at aicoderscope.com works on either. For the Ollama software side itself, aifoss.dev’s Ollama review covers setup beyond Apple hardware.

Sources

Last updated October 21, 2026. Prices and specs change; verify current rates before purchasing.

Recommended Gear

Mac Studio M5 Max 36GB — the fastest Mac under $3,000 for local AI: 460.8 GB/s, fits the whole 27–35B class at $2,499
Mac Mini M5 Pro 64GB — the only sub-$3,000 Mac that holds dense 70B Q4, in a silent always-on box
Mac Mini M6 32GB — cheapest new Apple hardware that clears the 27B bar, $1,299
Used RTX 3090 24GB — $1,456–$1,558; still 2–3× faster than any of the Macs above on everything it can hold

Was this article helpful?

Get the numbers before you buy

New GPU and mini-PC benchmarks, VRAM thresholds, and price checks — sent when there's something worth acting on, not on a schedule. No spam, unsubscribe anytime.