GMM-1 works through long documents, codebases and data — including requests over a million tokens — with higher accuracy, faster follow-ups and lower cost per task.
1M+ tokens · Up to 12× faster · 2× cheaper on measured long-document tasks
Sign up and you get GMM-1: a key and one line in Cline or Aider — a result in a couple of minutes, free on the 10,000 credits every new account starts with.
10,000 credits to start · 1 credit per 1,000 tokens in, 5 per 1,000 out
Why teams switch the model
General models lose the middle of a long document and guess when the input is larger than they can take.
Today
…and every answer has to be checked by hand
With GitMir GMM-1
Answers over the whole input, checked on documents with known answers. Exact counts over tables and logs. With a named project, it knows your product’s rules and says so before a change would break one.
Standard chat API. Pay per token from your credits — no subscription.
The same input and the same task for GMM-1 and for a general model. An answer counts only when it matches the known answer exactly.
Measured 5 Oct 2026 · generated inputs with known answers · tokens as counted by the model’s provider
Experiment 002 · October 2, 2026 · generated specifications, exact lists only · See the full results →
Measured 5 Oct 2026 · generated codebases with known answers · 60, 200 and 500 functions: GMM-1 10 of 10 each, a general model 10, 9 and 8
Measured 5 Oct 2026 · generated tables, JSON and logs with known answers
Measured 4 Oct 2026 · generated specifications and contracts with known results
Measured 4 Oct 2026 · the public models of cal.com, Supabase and Codex
Who it is for
Developer
And with a named project, a model that knows what must not break.
10 of 10 vs 8 of 10codebase tasks against a general model reading the same code
Analyst
Counted over the data you send — not numbers that only look plausible.
7 of 7 vs 0 of 2questions over a table, against a general model
Legal · product
What did not change stays word for word, and the contradictions are listed at the end.
Whoever pays
Billed per token from the team’s shared credits, with no separate subscription.
Measured on generated inputs with known answers; the full results are on this page.
“The model loses the middle of a long document”
Answers over the whole input, checked on documents with known answers.
“The input is bigger than the model can take”
It works with inputs larger than ordinary models accept.
“A confident answer with the wrong numbers”
Exact counts over tables and logs.
“The agent doesn’t know our product’s rules”
Name a project: the answer knows the product and warns before breaking a rule.
“Changing a function breaks the places that call it”
The change comes back with the places that call it.
“Running big documents through a top model is expensive”
A lower price per task, billed per token.
Make a key, then give your agent three things: the address, the key and the model name. Streaming, tool calls and JSON output work as usual.
Open Cline’s settings (the gear icon), choose the API provider “OpenAI Compatible” and fill in three fields.
https://app.gitmir.com/v1the key you made in your GitMir account — it starts with gmk_
gmm-1Press Done. Tasks, tool calls and streaming work as before.
Add GMM-1 to the models in Continue’s config.yaml.
models:
- name: GitMir GMM-1
provider: openai
model: gmm-1
apiBase: https://app.gitmir.com/v1
apiKey: gmk_…Pick GitMir GMM-1 in the model list.
Set the address and the key, then start Aider on GMM-1 from your repository.
export OPENAI_API_BASE=https://app.gitmir.com/v1
export OPENAI_API_KEY=gmk_…
aider --model openai/gmm-1Add GitMir as a provider in opencode.json and keep the key in an environment variable.
{
"$schema": "https://opencode.ai/config.json",
"provider": {
"gitmir": {
"npm": "@ai-sdk/openai-compatible",
"name": "GitMir",
"options": {"baseURL": "https://app.gitmir.com/v1", "apiKey": "{env:GITMIR_API_KEY}"},
"models": {"gmm-1": {"name": "GMM-1"}}
}
}
}Run /models and pick GMM-1.
Any client of the standard chat API works. With the openai package:
from openai import OpenAI
client = OpenAI(base_url="https://app.gitmir.com/v1", api_key="gmk_…")
r = client.chat.completions.create(
model="gmm-1",
messages=[{"role": "user", "content": "<your specification>\n\nIf field X changes, which requirements are affected?"}],
)
print(r.choices[0].message.content)The same call from Node.js, with the key in an environment variable.
import OpenAI from "openai";
const client = new OpenAI({baseURL: "https://app.gitmir.com/v1", apiKey: process.env.GITMIR_API_KEY});
const r = await client.chat.completions.create({
model: "gmm-1",
messages: [{role: "user", content: "<your specification>\n\nIf field X changes, which requirements are affected?"}],
});
console.log(r.choices[0].message.content);Use GMM-1 as the chat model of your own agent.
from langchain_openai import ChatOpenAI
llm = ChatOpenAI(model="gmm-1", base_url="https://app.gitmir.com/v1", api_key="gmk_…")
print(llm.invoke("<your specification>\n\nIf field X changes, which requirements are affected?").content)One request from the terminal — the quickest way to check a key.
curl https://app.gitmir.com/v1/chat/completions \
-H "Authorization: Bearer gmk_…" -H "Content-Type: application/json" \
-d '{"model": "gmm-1", "messages": [{"role": "user", "content": "Hello"}]}'Any other tool that lets you set the address of a chat API connects the same way: the address, the key and the model name.
The team that owns the key pays from its GitMir credits. No subscription for the model.
1 credit per 1,000 tokens you send
5 credits per 1,000 tokens
1 credit a request
A short question: 1 credit. In our tests a first question about a 140-requirement specification charged about 220 credits, and a follow-up about the same document about 11
10,000 credits with a new account — the same credits as every GitMir product
GitMir’s own model: more correct answers about long specifications, rule sets and documents — faster, at a lower price per task, over the standard chat API.
On generated documents, codebases, tables and logs whose correct answers are known in advance. GMM-1 and a general model got the same input and the same task, and an answer counted only when it matched the known answer exactly. Each benchmark names its size and date; Experiment 002 is published in full on the benchmarks page.
Any tool that lets you set the address of a chat API: Cline, Continue, Aider, OpenCode, the openai libraries for Python and JavaScript, LangChain, or plain HTTP. You give it the address https://app.gitmir.com/v1, a key from your GitMir account and the model name gmm-1.
GMM-1 answers knowing your product, and tells you before it starts when a request goes against a rule of your product. Name the project by its address, or set project in your client to its name or id — GitMir GMM-1 in your account lists your projects.
1 credit per 1,000 tokens you send and 5 credits per 1,000 tokens of the answer and the model’s work, at least 1 credit a request. A short question costs about 1 credit. A new account starts with 10,000 credits.
Requests and answers are not stored and not used for training. So that follow-up questions are faster, what GMM-1 reads — a document, pasted code or a data table — is kept in memory for up to 45 minutes after its last use and then erased. Files your agents write into a project that has no repository yet are kept only when an Admin turns on “Keep what agents write” for that project: they stay until you delete them, one by one or all at once, and reading them into the project's model uses credits.
A short question: about a second. A first question about a large document: one to three minutes, with progress streamed while it works; follow-up questions about the same document are several times faster. Tables, logs and JSON of more than a million tokens answer in under a minute; a first question about a codebase or a document that large can take up to 15 minutes.
Yes. Paste the files and ask: who calls a function, what a change reaches, or ask for the change itself — you get the new code of the function and of every part that calls it. On generated codebases of 60, 200 and 500 functions GMM-1 was right on 10 of 10 tasks; a general model reading the same input was right on 10, 9 and 8, and took 7 to 14 times longer. GMM-1 works with what you put into the request; it does not build or run the code.
Yes. Paste a CSV, a JSON export or a log and ask for counts, totals, averages, top lists or joins. The answers were exact in our tests at 300 rows and at 20,000 rows, in under a minute.
Yes. Send the document and one task: a table or JSON of everything it holds, a new version with its own corrections applied, a note for a team, a translation, or a contract with its contradictions settled by the contract’s own order of precedence. What you did not ask to change stays word for word, and contradictions that no rule settles are left as they are and listed for you.
10,000 credits with a new account.