LM15 Python

Install

```bash python3 -m pip install lm15

Quickstart

import os

from lm15 import Config, Message, OpenAILM, Request

lm = OpenAILM(api_key=os.environ["OPENAI_API_KEY"])

response = lm.complete(
    Request(
        model="gpt-4.1-mini",
        system="You are terse.",
        messages=(Message.user("Say hello in three words."),),
        config=Config(max_tokens=50, temperature=0.2),
    )
)

print(response.text)
print(response.finish_reason)
print(response.usage.total_tokens)

Hello there, friend.
stop
27

The mental model is one straight line:

Message parts → Message → Request → ProviderLM → Response
                              │
                              └── stream() → StreamEvent → materialized Response

Optional extra for websocket live sessions:

python3 -m pip install 'lm15[live]'

Or from source, for development:

lm15 has zero required dependencies — it is stdlib-only, including its HTTP transports.

Media and non-chat endpoints

Multimodal input uses typed media parts (ImagePart, AudioPart, DocumentPart, ...):

import os

from lm15 import ImagePart, Message, OpenAILM, Request, TextPart

lm = OpenAILM(api_key=os.environ["OPENAI_API_KEY"])

request = Request(
    model="gpt-4.1-mini",
    messages=(
        Message.user([
            TextPart("Describe this image in a few words."),
            ImagePart(
                url="https://raw.githubusercontent.com/github/explore/main/topics/react/react.png",
                media_type="image/png",
                detail="low",
            ),
        ]),
    ),
)

print(lm.complete(request).text)

This image shows a blue atomic symbol, often used to represent an atom or atomic energy.

Non-chat endpoints have separate request/response types — EmbeddingRequest, ImageGenerationRequest, AudioGenerationRequest, FileUploadRequest, BatchRequest, LiveConfig:

from lm15 import EmbeddingRequest

embeddings = lm.embeddings(
    EmbeddingRequest(
        model="text-embedding-3-small",
        inputs=("hello", "world"),
    )
)
print(len(embeddings.vectors), len(embeddings.vectors[0]))

2 1536