Models

Model types

Which models you call over HTTP and which run in your own program, and what each takes.

Every Myles model is a decision model. It reads a situation, a question and a list of options, and returns a probability for every option in one forward pass. No Myles model writes text, and we do not offer a text or chat model: there is no /v1/chat/completions or /v1/completions endpoint, and a request to either gets 404. For drafting, summarising or conversation, use a general-purpose LLM beside Myles (see Myles vs general LLMs).

What differs between models is how you call them. There are two kinds.

GPU decision models: the Decision API

Myles-G is served only for decisions: the server reads one answer position and never returns generated text. Myles-G is called over HTTP, on our hosted endpoint (Myles-G 12B and Myles-G 31B) or on a server you run yourself, with the same request.

• Endpoint: POST /v1/decide. POST /v1/systemone is the same, for existing decision-model clients.
• Request: state (text or JSON), questions (named noul, choice or score questions), an optional model, and, on Myles-G 12B and Myles-G 31B, optional images in assets (see Images).
• Response: one JSON object, sent when every question is answered. There is no streaming.
• Limits: 16 questions and 26 options per question; 32,768 tokens of context on the hosted Myles-G 12B, 16,384 on the hosted Myles-G 31B. See Errors and limits.

curl https://api.launchables.ai/v1/decide \
  -H "Authorization: Bearer $LAUNCHABLES_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "state": "Your parcel is on hold. Pay the £1.99 customs fee here to release it today.",
    "questions": {
      "scam": {"type": "noul", "instructions": "Is this message a scam, phishing or fraud attempt?"}
    }
  }'

{
  "model": "myles-g-12b",
  "answers": { "scam": { "type": "noul", "noul": 0.97 } },
  "usage": { "input_tokens": 104, "output_tokens": 1 }
}

Illustrative probability and token counts. On Node, the JavaScript SDK sends the same request:

import { Launchables, noul } from "launchables";

const client = new Launchables(); // reads LAUNCHABLES_API_TOKEN
const resp = await client.decide(
  "Your parcel is on hold. Pay the £1.99 customs fee here to release it today.",
  { scam: noul("Is this message a scam, phishing or fraud attempt?") },
);
console.log(resp.answers.scam.probability);

Every field is in the API reference, and each question type in Question types.

CPU decision models: in your own program

Myles-400M, 150M and 22M are small encoders. We do not host them and they have no HTTP API: you run the model file on your own CPU, with the short Python helper (myles.py) that comes with the weights. The API keys, the console and the SDK are not used.

• Call: decide(state, question, options), one question per call.
• Options: a list of at least two short texts. There are no question types: for yes or no, pass ["no", "yes"]; for a choice, the labels; for a scale, the levels in order.
• Returns: a dictionary of option to probability.
• Input: question, options and state together, up to 4,096 tokens on 400M, 1,536 on 150M and 320 on 22M. Longer states are cut.

from myles import Myles   # myles.py, in the folder with the weights

model = Myles(folder, threads=4)
model.decide(
    state="Your parcel is on hold. Pay the £1.99 customs fee here to release it today.",
    question="Is this message a scam, phishing or fraud attempt?",
    options=["no", "yes"],
)
# {'no': 0.1032, 'yes': 0.8968}

Output from Myles-150M. The weights are on Hugging Face for internal evaluation, on request: Myles-400M, Myles-150M and Myles-22M. Production use is under a licence. Myles-22M also runs in a browser: try it.

Which to use

• Several questions about one situation, or the hardest decisions: Myles-G, through the Decision API.
• No GPU, no network, or text that must stay on the device: a CPU model, in your own program.
• Writing or summarising text: not a Myles model. Use a general-purpose LLM, and Myles for the decisions around it.