GMM Local is GitMir’s 30B model for documents, code and data. It works through inputs far beyond its own context window while running entirely inside your infrastructure.
30B model · NVIDIA 24 GB+ · Apple Silicon 32 GB+ · Standard chat API
$29 per machine / month · NVIDIA 24 GB+ or Apple Silicon 32 GB+
Powered by GMM-30B.
Why teams run it locally
Some work can’t go to a cloud model — and most local models give up on long documents.
Today
…so the work is done by hand
With GitMir GMM Local
It reads files and folders by path, works through inputs far beyond its context window and writes the answer back into a document. One machine can serve everyone on your network.
$29 per machine a month. No per-token charges.
The same model, the same input and the same task — once on its own, once as GMM-30B. An answer counts only when it matches the known answer exactly.
Measured 4–5 Oct 2026 · generated specifications with known answers · exact lists only
Measured 4 Oct 2026 · one message with the document and thirteen tasks
Measured 4 Oct 2026 · documents written for the test, answers known in advance
Measured with the model at full precision; GMM Local runs the same model compressed to fit a single machine.
Who it is for
Whoever can’t share the data
Renewing the license never sends what you ask or the files you give it.
Whoever counts the money
$29 a month for the machine that runs the model — however long the documents are.
Whoever works with long documents
Specifications, contracts and code far larger than the model’s window — on one machine.
2,000 requirementsAll 16 requirements in a dependency chain returned correctly
The team
Only the machine that runs the model counts. Computers that use it over the network do not.
“We can’t send documents to the cloud”
The model and the GitMir reasoning run on your own machine.
“On large documents we pay for every token”
A price per machine, with no per-token charges.
“Local models can’t handle long documents”
Inputs far larger than the window, answered in full.
“The model won’t open my files”
Files and folders are read by their path in the message.
“The answer still has to be copied into a document”
It writes .docx, .md, .csv and other formats itself.
“There’s no internet where we work”
It keeps working offline; the license renews when the machine is online, at least once a week.
A free check tells you whether this machine runs GMM Local before you pay. Then your GitMir account gives you one command with your license already in it: it switches the machine on and starts it — nothing else to type.
curl -fsSL https://gitmir.com/install | shIt downloads the small GMM Local program — not the model — and checks memory, the GPU and free disk space. It says whether the machine runs GMM-30B, runs it slowly, or cannot run it, and why. Nothing is switched on and nothing is charged.
$29 per machine per month. Get GMM Local →
curl -fsSL https://gitmir.com/install | sh -s -- --token gmmt_…It is on the GMM Local screen as soon as you subscribe. It downloads GMM-30B, about 18 GB, once — step by step, in percent — starts it and says when it answers.
Talk to it in the terminal: gmm chat. Or point your tools at
http://localhost:7331/v1Model gmm-30b. To update, run gmm update.
macOS on Apple Silicon and Linux · NVIDIA 24 GB+ · Apple Silicon 32 GB+
Your editor, your agents and your own code connect to it like to any hosted model: the address and the model name.
http://localhost:7331/v1gmm-30bNone on this machine. With gmm run --lan, other machines use the key from gmm key.
from openai import OpenAI
client = OpenAI(base_url="http://localhost:7331/v1", api_key="local")
r = client.chat.completions.create(
model="gmm-30b",
messages=[{"role": "user", "content": open("contract.md").read() + "\n\nWhich clauses contradict each other?"}],
)
print(r.choices[0].message.content)curl http://localhost:7331/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model": "gmm-30b", "messages": [{"role": "user", "content": "Hello"}]}'Provider: OpenAI Compatible
Base URL: http://localhost:7331/v1
API key: local (any text: GMM Local on this machine needs no key)
Model: gmm-30bgmm ask "Which requirements depend on REQ-1041?" specification.mdThe standard chat API — /v1/chat/completions and /v1/responses, with streaming and tool calls. Any tool that lets you set the address of a chat API connects the same way.
The model, GitMir reasoning and the API on your computer.
Works offline between weekly license renewals. Your data never leaves the machine.
gmm runOne workstation serves every developer and agent on your local network.
Only the machine that runs the model counts.
gmm run --lanGMM on your own servers, inside your own network.
An annual agreement — talk to us.
Talk to usThe free check (step 1 above) tells you whether a machine runs it before you subscribe and before the model is downloaded; on an installed machine, gmm check says the same.
A GPU with 24 GB of memory or more — an RTX 4090 or 5090 class card. Linux.
M1 or later with 32 GB of memory at least; 64 GB is comfortable. macOS.
About 20 GB free for the model and the model server.
It runs on the processor too, with 32 GB of memory — slowly.
Not supported yet.
The model runs on your hardware, so there is nothing to meter.
$29 per machine per month
The machine that runs the model. Computers on your network that use it do not.
Stays on the machine. GMM Local talks to GitMir only to renew its license — the key, a fingerprint of the machine, its system and the app version, never your inputs. What it learns about a document stays on the machine for 72 hours; gmm forget clears it.
It keeps working without the internet; the license renews itself the next time the machine is online (at least once a week).
Any time, in your GitMir account. The machine works until the end of the paid month.
GitMir’s 30B model with GitMir reasoning, running on your own hardware: it works through inputs far beyond its context window, and your data never leaves your computer. Any tool that speaks the standard chat API connects to it.
Subscribe in your GitMir account. The GMM Local screen then shows one command with your license already in it: paste it into a terminal on the machine. It installs GMM Local, switches the machine on, downloads the model (about 18 GB, once) step by step in percent, starts it and says when it answers. Then talk to it with gmm chat, or connect your tools.
No. The model and GitMir reasoning run on your machine, and your inputs and answers stay there. GMM Local talks to GitMir only to renew its license: it sends the license key, a fingerprint of the machine, its operating system and the app version — never what you ask. What it learns about a document stays on your machine for 72 hours; gmm forget clears it.
An NVIDIA GPU with 24 GB of memory or more on Linux, or a Mac with Apple Silicon and 32 GB of memory or more; about 20 GB of free disk. It also runs on the processor with 32 GB of memory, slowly. Run gmm doctor to see what your machine can do before anything large is downloaded.
Whole documents, codebases and data — far more than a model takes in at once. It is answered in full, not cut to fit: in our tests a 2,000-requirement specification was answered exactly, all 16 requirements of a chain.
Anything that lets you set the address of a chat API: Cline, Continue, Aider, OpenCode, Codex, the openai libraries for Python and JavaScript, or your own code. The address is http://localhost:7331/v1 and the model is gmm-30b; /v1/chat/completions and /v1/responses both work.
Yes. gmm run --lan opens it to your local network, and the other computers use it with the key from gmm key. Only the machine that runs the model counts toward the subscription.
Yes, once it is installed and the model is downloaded. The license renews itself whenever the machine is online; it needs to be online at least once a week.
$29 per machine per month, cancel any time. There is no charge per token: the model runs on your hardware.
Use GitMir GMM-1 in the cloud: the same family of models over an API, paid per token from your GitMir credits.
$29 per machine per month.