Gemma 4 just got faster.Up to 90% faster on your Mac.
Ollama turned on multi-token prediction for Gemma 4 through Apple's MLX engine — and it's a big jump. I ran the real numbers on my M4 Max, then had Gemma 4 build a dragon realm, a landing page and a Skyrim world. All on my own machine. All free. Here's the data and the actual builds.
Ollama says Gemma 4 is ~90% faster with MLX
This is the post that kicked it off. Ollama turned on multi-token prediction (MTP) for Gemma 4 by default, powered by Apple's MLX engine. The clever bit: it auto-tunes how many tokens to guess ahead, so it never slows you down when guessing stops helping. Below, I test whether that 90% is real.
The official sources. Read it, run it yourself.
"Gemma 4 is now nearly 90% faster on Apple Silicon with Ollama using MLX… Ollama automatically tunes how many tokens to draft as it runs, so it never slows generation down."
— @ollama, 1 July 2026
I stopped paying to build small stuff.
Before
Every little build — a landing page, a quick game, a test scene — went through a paid cloud model.
It was fine, but it added up, and it meant waiting on someone else's servers.
My local models were free, but slow enough that I only reached for them now and then.
So the free option sat mostly unused.
Then Gemma 4 got ~60% faster on my Mac overnight — and on coding, closer to 90%.
After
Now the everyday building happens on my own machine, at , at a speed that finally feels good.
I ask, it builds, I iterate — no meter running.
The cloud is for the truly hard stuff. The daily 90% runs local and free.
Same laptop you already own. A model that's now genuinely fast. Zero per-build cost.
Why listen to me on this.
I run an AI agency with 70+ people where AI handles about 80% of the ops, and a room of operators doing the same across every kind of business.
No invented quotes here. The wins are real and written by the members themselves — agency owners, ecom founders, creators, solo operators across 38 countries. Read them in their own words.
Read the 158-page wins doc →Commit to running one build local tonight.
You've seen the claim. Below is the real data and the real builds.
Here's the deal I want to make with you.
Before you sleep tonight, update Ollama, pull the MLX build, and make Gemma 4 build you one thing — free, on your own machine. Just one. A landing page, a little game, whatever.
Because the people who move their everyday building to a fast, free local model stop thinking about cost entirely — and they iterate more, so they ship more.
Be one of them.
Commit to the transition. One local build, tonight. That's how it becomes your default.
Is it actually 90% faster? On coding — yes.
I ran the same prompts through the old GGUF build (llama.cpp) and the new MLX build (with MTP) on my M4 Max. Here's the raw tokens-per-second — higher is faster.
The honest read: on general chat it's about +60% (39 → 62 tok/s). On coding — the workload the 90% claim is measured on — it climbs to ~90%+, and on very repetitive code I saw it hit 2.6× (27 → 71 tok/s). MTP guesses several tokens ahead, and the more predictable the code, the more it gets right — so it speeds up exactly when you're building.
No — that's the smart part. When Gemma 4 drafts a token that turns out wrong, it's simply thrown away, so the output is identical to before — just produced faster. And Ollama auto-tunes how far ahead it guesses, backing off to normal speed when guessing stops helping. You get the speed with zero quality trade-off.
Two lines to turn it on.
The catch most people miss: the speedup only applies to the MLX build. If you're on a GGUF Gemma 4, you get none of it. So pull the MLX one:
ollama --version # need 0.31+
ollama pull gemma4:12b-mlxThat's it — MTP is on by default. One speed tip from the thread: set keep_alive: -1 so the model never unloads between calls (no cold-load wait). On an M4 Max the 12B build lands around 62 tok/s general, 70+ on code, running 100% on the GPU.
Run fast Gemma 4 inside the Agent OS.
The speedup is one model getting quicker. The Agent OS is where that fast, free local model becomes your everyday engine — wired into one dashboard with everything else.
You're not buying a tool. You're getting the whole operating system I run a seven-figure business on.
Get the Agent OS →No — and this update is exactly why. The Agent OS runs the everyday 90% on a free local model — now a fast one — on your own machine, so most work costs $0 and never leaves your Mac. Free APIs slot in for more, and for the frontier stuff it drives the CLIs you already pay for (your Claude subscription already includes the Claude CLI — the Agent OS plugs straight into it, so you're not paying twice).
It's a layer on top of what you already own, not a new meter. And inside the Boardroom there are full token-optimisation tutorials so you cut usage to the bone.
So what can fast, free Gemma 4 actually make?
Here's the honest bit. The MLX speedup is the base Gemma 4 getting quicker — great for chat, agents and iterating. For building rich stuff, I use the coder-tuned Gemma 4 (still Gemma-4 12B, still local, still free). Every build below was made by Gemma 4 on my machine — nothing hand-drawn by me. Click any one to open the live build.
The dragon realm is a full walkable snow world — mountains, snow-capped pines, falling snow, a first-person sword. The landing page has a real nav, hero, features and pricing. The forest is a low-poly open world you can walk around. All Gemma 4, all local, all . (Live 3D needs a real browser to render; if a preview looks flat, open it fullscreen.)
How most people build small stuff. And how I do it now.
- Every landing page + test scene hits a paid API
- Wait on someone else's servers
- Local model too slow to bother with
- Iterate less because each try costs
- Your work leaves your machine
- Everyday builds run on your own Mac, free
- ~60-90% faster than last week — feels instant
- Iterate all you want, no meter
- Cloud saved for the genuinely hard jobs
- Nothing leaves your machine
What's holding you back.
Wrong: "Local models are too slow to be useful."
Right: That was true a week ago. On coding, MLX Gemma 4 is now ~90% faster and lands 70+ tokens a second on a laptop — quick enough to be your daily driver, not a novelty.
Wrong: "Free means worse."
Right: The speedup changes nothing about the output — wrong guesses are thrown away, so it's identical, just faster. And the builds above are real, from a free local model.
Wrong: "This is for coders with big rigs."
Right: It's two commands on a Mac you already own — ollama pull gemma4:12b-mlx and go. If you can copy-paste one line, you can run it.
158 pages of members already running this local stack — real businesses, real wins, in their own words.
Read the 158-page wins doc →Fast + free is the whole point.
Gemma 4 just became fast enough to run your everyday AI for free, on your own machine. That's the piece that makes an Agent OS actually cheap to run all day.
Inside the AI Profit Boardroom you get the full Agent OS — the fast local engine, Free Claude Code on Gemma 4, Agent Kanban, the Claude Workspace, every CLI you already pay for in one dashboard, the AI Mastermind, memory that knows your business, a 30-day roadmap, daily tutorials, coaching calls, and 3,900+ founders across 38 countries building alongside you. Every new update — like this speedup — gets folded in the week it lands.
It's the operating system I run a seven-figure business on. You get the whole thing.
Get the Agent OS →Run ollama pull gemma4:12b-mlx tonight, and make it build you one thing. I'll see you in the next one.
