Muse Spark 1.3

Muse Spark is Meta’s proprietary line from Meta Superintelligence Labs — unlike Llama, its weights are not released. Version 1.3, from September 2026, is built for long agent sessions and coding across large repositories; Meta reports 75.4% on DeepSWE and 88.8% on Terminal-Bench 2.1.

Its most unusual strength is recall over long inputs: 98.1% on the MRCR test between half a million and a million tokens, so it finds what you need even at the far end of a huge document. Meta also says it is harder to trick with injected instructions, and more careful about actions that cannot be undone.

For scale: 1,000 requests of about 3,000 tokens in and 1,000 out cost $10.40 — a long prompt and a page of answer each. Every rate is published and stays the same from one request to the next. See every model.

Muse Spark 1.3 is billed in rupees, on the same monthly invoice as every other model your team uses — no foreign card, no separate Meta account, and one key for all of it.

Prompts sent to Muse Spark 1.3 are processed outside Sri Lanka. If a workload has to stay in the country, Qwen3.8-27B (Colombo) runs on our own servers in Colombo.

Use the OpenAI SDK you already have — point it at Roar AI and name the model. Read the docs.

from openai import OpenAI

client = OpenAI(base_url="https://api.roar-ai.com/v1", api_key="roar_live_…")

reply = client.chat.completions.create(
    model="muse-spark-1.3",
    messages=[{"role": "user", "content": "Summarise this contract in plain English."}],
)