Muse Spark for Developers: API, Muse Code and the Tooling Around It

Muse Spark is Meta's closed-weights reasoning model for coding and agent work, and developers reach it two ways: the Meta Model API at https://api.meta.ai/v1, which is drop-in compatible with the OpenAI and Anthropic SDKs, or Muse Code, Meta's terminal coding agent. The current version, Spark 1.3, has a 1,048,576-token context window and costs $1.25 per million input tokens on the standard tier. A community has already built editor bridges, delegate plugins and harness fixes around it.

Muse Spark versions and what changed

Date	Release	Source
Apr 8, 2026	Muse Spark announced; API in private preview	Meta
Jul 9, 2026	Spark 1.1 and the Meta Model API public preview	Meta
Aug 5, 2026	Spark 1.2 and Muse Code (beta)	Meta
Sep 2, 2026	Spark 1.3, with a max reasoning mode	Meta

Meta says 1.3 uses about 20% fewer tool calls and 25% fewer tokens than 1.2, and that bigger models and a "Muse Spark open weights release" are coming. Spark 1.3 also powers the consumer Muse agent.

Calling the Meta Model API

Per Meta's docs, auth is a Bearer key (MODEL_API_KEY), and the API accepts the Responses, Chat Completions and Messages formats. With the OpenAI Python SDK, you change the base URL and key:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key=os.environ["MODEL_API_KEY"],
)

resp = client.chat.completions.create(
    model="muse-spark-1.3",
    messages=[{"role": "user", "content": "Summarize this diff in three bullets: ..."}],
)
print(resp.choices[0].message.content)

Spark 1.3 supports parallel tool calls with streamed arguments, structured output, prompt caching, and text, image and video/document input. Other models on the same API include muse-image-1.0, muse-voice-transcribe-1.0 and sam-3.1.

One integration gotcha: Spark emits a non-standard SSE event that trips some OpenAI-compatible tools. muse-spark-anywhere packages fixes for pi, opencode and Agent Orchestrator.

Standard vs contributor tier

This choice matters more than the model version.

Tier	Input / Output per 1M	Rate limit per team	Data
Standard	$1.25 / $4.25	3,000 RPM, 4M TPM	Not used for training
Contributor	$0.10 / $0.20	100 RPM, 3M TPM	Meta may train on it

Source: Meta pricing. Contributor is roughly 12x cheaper on input and 21x cheaper on output, which is why so many community pipelines use it for bulk work. Keep proprietary code on Standard.

Installing Muse Code

Muse Code is the terminal coding agent co-trained with Spark. Install it with:

# macOS / Linux
curl -fsSL https://dev.meta.ai/install.sh | sh

# Windows (PowerShell)
irm https://dev.meta.ai/install.ps1 | iex

Run muse, then sign in through the browser or paste an API key. Meta lists subagents in isolated git worktrees, auditable and replayable event logs, and built-in skills like /plan, /grilling and /taste. Plans are $5, $15 and $50 a month. DataCamp reports Muse Code defaults to the contributor model, so check the setting.

For a sandboxed setup, Oleg Šelajev's Docker Sandboxes kit only allows network access to the Meta services Muse Code needs.

Using Spark in your editor

You don't have to live in the terminal. From the catalog:

• Zed and JetBrains: muse-acp is a dependency-free Rust bridge between the Agent Client Protocol and Muse Code's session protocol. muse-code-acp does the same in Node.
• VS Code: Muse Spark Code adds streaming chat, reviewable diffs, permission modes and watchable subagents.
• Desktop GUI: Helicon wraps the CLI with projects, history and diffs.
• Remote: Muse Code Remote drives your terminal agent from any browser without opening ports.
• Other harnesses: a pi extension, an OpenRouter adapter, and Cursor, Vercel AI Gateway and OpenCode support.

Guardrails exist too: Maestro adds command guardrails, secret redaction and cost telemetry, and the Agentic Control Plane plugin policy-checks every tool call.

Patterns that work: Spark as the cheap worker

The most common architecture in the catalog pairs Spark with a pricier planner or reviewer.

• sol-loop: GPT plans, Spark executes and returns evidence and diffs. The author says over 99% of tokens run on Muse.
• ocodex: contributor-tier workers fan out while one paid supervisor audits every claim.
• delegate-to-muse: a Claude Code skill that hands tasks to muse exec and then re-runs tests instead of trusting the summary.
• muse-spark-crew: ten specialist subagents with review gates.

The shared lesson: verify Spark's output mechanically. Tests, diffs and ledgers do the trust work. See Muse vs Claude for more on mixing the two.

What people have built with Spark

• A single-file Three.js platformer designed, coded and play-tested in one autonomous session, passing 25/25 browser checks.
• A minigolf game converted to Quest 3 VR with Muse Code and the Unity CLI, plans and prompts published.
• Tarquin, a voice robot butler on an M5Stack StackChan using Spark 1.3 via the API.
• A Windows computer-use agent in native C++ driving Spark through the Responses API.
• An OS, compiler and CPU in Rust and Verilog.

Browse more under tools and apps and GitHub builds.

How good is it, honestly?

Meta reports 75.4 on DeepSWE v1.1 and 88.8 on Terminal-Bench 2.1. Independent results are mixed: eesel says it trails Opus 5 on four of six agent evals, Morgan Linton saw serious issues at lower effort levels, and MindStudio's game test disappointed. Paweł Huryn's 105-bug test had Spark 1.3 at max fix 33, tying Fable 5.1 (high) and ahead of Opus 5 at 27, as reported by the author. Use high or max effort for real work, and benchmark on your own repo.

To learn it end to end, freeCodeCamp has a three-hour course ending in a Go, SQLite and Docker build.

Frequently asked questions

Is the Muse Spark API compatible with the OpenAI SDK?

Yes. Meta says the Meta Model API is drop-in compatible with the OpenAI SDK, the Anthropic SDK and OpenAI-compatible CLIs. Point the base URL at https://api.meta.ai/v1 and use your MODEL_API_KEY.

What is the difference between Muse Spark and Muse Code?

Muse Spark is the model. Muse Code is Meta's terminal coding agent that runs Spark, with subagents in git worktrees, event logs and built-in skills.

What is the Muse Spark contributor tier?

It's a much cheaper pricing tier ($0.10 input and $0.20 output per million tokens) where Meta may train on your data. The standard tier costs more but prompts are not used for training.

Is Muse Spark open source?

No. Spark is closed weights. Meta has said a Muse Spark open-weights release is coming, and its current open model is Muse Glimmer 30B under Apache 2.0.