Global Agent

Global Agent is a runtime that lets you run massive language models locally on an Apple silicon Mac. It keeps the core of a model in memory and streams the rest from your drive as needed, so models in the 80B to 400B+ range run on a 16 to 48 GB machine. Nothing runs in the cloud, so your code and conversations stay with you.

Streaming experts is of course not a new or novel idea. What's new is a lot of little things built on top of that. They fight quality loss, speed up prefill and decode even more, and add up to a runtime that's actually efficient enough to be worth using. That's what makes the benchmarks below possible. Plenty of them are still experimental.

It ships with everything you need to use those models:

• Built-in IDE. An editor with a side panel, an agent that reads, searches, edits and creates files and runs commands, and diff review for every change.
• CLI. The ga command gives you the same agent in your terminal.
• Local server. Run it headless and connect your own tools and integrations.

Download: globalagent.dev

This repo only holds the overview and the benchmark numbers. There is no source code here.

What you get

• Massive models on a laptop, without a cloud account or a token meter.
• One agent across the editor, the terminal and the server.
• Every edit shows up as a diff you can keep or undo. Commands need your approval.
• Works offline once a model is downloaded.

Requirements

• Mac with Apple silicon
• macOS 13 or newer
• 16 GB of RAM or more
• Free disk space for the model you pick (see the sizes below)

Faster internal storage helps a lot with the biggest models.

Benchmarks

Every number below comes from Global Agent 1.2.0 doing the same agent job on each model: "Build a simple express app with a ping/pong endpoint, and a web ui that calls it and displays the response. Start both the server and ui when complete." The agent plans it, writes package.json, server.js and the page, runs npm install, starts the server, and reports back, then a second chat starts. The prompts and tool schemas are the real ones the app sends.

• Machine: Apple M5 Pro with 48 GB. Smaller RAM sizes are simulated by limiting what the app is allowed to use, so a real Mac with the same RAM and a slower drive or chip will be slower.
• Exact: every expert the model picks is used. The output is the model's own.
• Speed: the speed slider at 20, the same on every model. Much faster, because it can skip experts that are not ready in memory, while tool calls and the start of every turn are held to a stricter setting. Top-1 agree shows how often the model's top token still matched exact mode (1.00 means no difference).
• Decode tok/s: median tokens per second across the task's requests (so one hot step does not skew it).
• First turn: time to the first token on a cold chat, with nothing cached.
• Full task: wall time for the whole task, from the first request to the new chat's reply.
• Peak GB: peak memory the process used.

Treat these as a guide and not a promise. Your speed will depend on your Mac, your drive, and what else is running.

Qwen3-Coder-Next 80B-A3B (6-bit)

80B total, 3B active, 65 GB on disk.

RAM	Mode	Decode tok/s	First turn (s)	Full task (s)	Peak GB	Top-1 agree
16 GB	exact	6.44	17.4	129	13.3	1.00
16 GB	speed	41.2	17.4	57	13.3	0.91
32 GB	exact	13.9	16.1	77	19.9	1.00
32 GB	speed	34.9	15.9	47	19.9	1.00
48 GB	exact	17.7	16.2	57	31.5	1.00
48 GB	speed	34.5	15.8	44	31.5	1.00

Qwen3.8-Flash-Next 229B (6-bit)

229B total, 7B active, 148 GB on disk.

RAM	Mode	Decode tok/s	First turn (s)	Full task (s)	Peak GB	Top-1 agree
16 GB	exact	3.53	38.2	373	18.7	1.00
16 GB	speed	17.1	37.3	128	18.7	0.30
32 GB	exact	6.18	31.7	233	20.6	1.00
32 GB	speed	18.3	31.5	116	20.6	1.00
48 GB	exact	7.88	32.0	186	31.5	1.00
48 GB	speed	19.2	32.2	112	31.5	1.00

Qwen3-235B-A22B (4-bit)

235B total, 22B active, 132 GB on disk.

RAM	Mode	Decode tok/s	First turn (s)	Full task (s)	Peak GB	Top-1 agree
16 GB	exact	1.13	46.6	605	11.3	1.00
16 GB	speed	3.46	45.9	563	11.3	0.19
32 GB	exact	1.82	46.6	464	20.9	1.00
32 GB	speed	9.71	44.6	183	20.9	0.88
48 GB	exact	2.62	47.6	343	32.6	1.00
48 GB	speed	9.19	42.1	149	32.6	0.98

MiniMax-M3 426B (4-bit)

426B total, 24B active, 240 GB on disk.

RAM	Mode	Decode tok/s	First turn (s)	Full task (s)	Peak GB	Top-1 agree
32 GB	exact	1.41	44.7	808	21.2	1.00
32 GB	speed	3.53	43.7	366	21.2	0.38
48 GB	exact	2.20	43.1	526	31.7	1.00
48 GB	speed	4.62	44.9	295	31.8	0.75

MiniMax-M3 needs 32 GB or more: on 16 GB its backbone alone leaves no room for the expert cache, so the app does not load it there.

Raw data is in benchmarks/summary.json.

Choosing a model

Bigger models are smarter and much slower on small Macs. As a rough guide:

• 16 GB: the 80B model is the practical choice, and it is fast. The bigger models run in exact mode if you are patient.
• 32 GB: the 80B model flies, the 229B model is comfortable on speed with no drift, and the 235B model is usable.
• 48 GB: the 229B and 235B models are a good fit on speed, and the 426B model works. On the 426B, use exact when you want its exact output.

Install

1. Download the disk image from globalagent.dev.
2. Open it and drag Global Agent into Applications.
3. Launch it once. The first launch installs the ga command.

Full setup steps and docs are on the website.

Heads up

Running large models pushes a Mac hard. Expect heat, fan noise, heavy memory and drive use, and a slow machine while a big model is working. Keep backups and review diffs before you keep them. The full terms and disclaimer are on the website.

Links

• Website and download: globalagent.dev