Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Run Muse Spark in another harness via the Model API

Written November 9, 2026

Meta Model API serves Muse Spark 1.3 at an OpenAI-compatible endpoint, with a cheap Contributor tier. The IDs, the prices, and the swap nobody has tested.

The swap, as the docs describe it

Meta's Models page lists Muse Spark as the company's model for agentic and coding work at scale — multi-step tool loops, software engineering assistants, subagent tasks, and long-context reasoning — and names muse-spark-1.3 as the version recommended for new work, alongside muse-spark-1.2 and muse-spark-1.1. All three take text, image, video, audio, and PDF input and return text, with a 1,048,576-token context window. The API itself lives at https://api.meta.ai/v1 and accepts the Responses, Chat Completions, and Messages formats, so any agent tool that speaks an OpenAI-compatible endpoint can in principle point at it: set the base URL to https://api.meta.ai/v1, authenticate with a MODEL_API_KEY bearer token, and request muse-spark-1.3. Meta's own example does exactly that with the OpenAI client, calling client.models.list() against the Meta base URL.

That is the whole recipe the docs give, and it is also where the docs stop. No swap test exists. No agent tool was pointed at this endpoint for this page, no session was run, and no tool loop was driven through it. Everything below is what Meta's documentation states, not what a run proved. If you try the swap, the first thing to check is whether your tool actually completes a tool call against the endpoint — an OpenAI-compatible shape is not a guarantee that every tool's loop survives the trip.

Two tiers, and the one that trains on your prompts

Meta's pricing page splits Muse Spark into two modes, and the split is the decision that matters most. Standard mode costs $1.25 per million input tokens, $4.25 per million output tokens, and $0.15 per million cached input tokens — and Meta states plainly that Standard-mode content is not used to train Meta models. Contributor mode runs the same models at $0.10 input, $0.20 output, and $0.002 cached input per million tokens, in exchange for permission to use your prompts, code, files, context, and the model's responses to train future Meta models. You pick Contributor mode by calling a different model ID: muse-spark-1.3-contributor or muse-spark-1.2-contributor. There is no Contributor variant of 1.1, and Contributor mode is not available for Muse Wonder on Model API at all.

The rate limits split the same way, and they apply per team, not per API key — every key in a team counts toward the same budget. Standard Spark 1.3 allows 3,000 requests per minute and 4,000,000 tokens per minute; the Contributor variant allows 150 requests per minute and 3,000,000 tokens per minute. A long agent session on the cheap tier can hit the request cap long before it hits the token cap, so the tier that looks twelve times cheaper on input tokens is also twenty times tighter on requests.

What not to reach for

Two nearby Muse things are not reachable through this API, and the docs say so. Muse Wonder Full — the higher-capability Wonder variant — runs in Muse Code only, not on Model API; the plain muse-wonder-1.0 is the one the API serves, at $4.00 input and $20.00 output per million tokens, Standard mode only. And Muse Spark Open, the 816B-parameter open-weight release, is not callable over Model API at all: you download the weights and serve them yourself, or call them through a hosted inference provider. If a tool offers you a "Muse Spark" model through some other route, check which of the three — hosted Spark, Wonder, or open weights — it is actually serving before you compare prices.

Verdict

Start on Standard mode with muse-spark-1.3 and the documented base URL, and treat the first tool-using session as the test, not the rollout. Contributor mode is the right price only for prompts you would not mind Meta training on — prototype code, public examples, throwaway experiments — never for proprietary code or anything confidential, because the discount is paid for with your content. Until someone publishes a swap test for your setup, assume the endpoint shape works and the tool loop is unverified.

What can go wrong

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Judge Muse Spark Open before you download it: 816B weights, a custom license, and a cluster-sized hardware bar · Model prices: what the major models cost · Check whether your session is really on Muse Wonder or Wonder Full — and which route you are paying through