# Muse Spark Open, Explained

> Independent explainer for Meta's Muse Spark Open open-weight model. Not affiliated with Meta. Updated 2026-10-16. Full page: https://lab.awaara.cn/muse-spark-open/

## What it is

Muse Spark Open is Meta's open-weight AI model, released October 15, 2026. It has 816 billion total parameters with 42 billion active per token (mixture-of-experts, 256 experts, top-8 routing), a 1-million-token context window, and accepts text, images, video and audio as input. Output is text. Meta built it for agentic coding: reading codebases, debugging, and calling tools over long sessions.

Unlike Muse Wonder (Meta's closed frontier model, API-only at $4/$20 per 1M tokens), Spark Open's weights can be downloaded and run on your own hardware.

## Specs

- Released: October 15, 2026, by Meta
- Architecture: MoE transformer with perception encoder; 256 experts, 8 active per token
- Parameters: 816B total / 42B active per token
- Context: 1M tokens
- Checkpoints: BF16 (~1.5 TB), NVFP4 4-bit (~0.41 TB)
- License: Muse Model License — limited commercial terms, not unrestricted open source
- Base: started from Muse Spark 1.2 (August 2026) plus safety post-training
- Frameworks: llama.cpp, vLLM, SGLang recipes; Hugging Face hosted inference endpoint (OpenAI-compatible API)

## The Mac story

Meta partnered with Apple to tune the model for Apple's MLX framework. The demo ran on two Apple M5 Ultra Mac Studios behaving as one system with about 1 TB of combined unified memory (each M5 Ultra GPU addresses up to 512 GB; 1.2 TB/s memory bandwidth). Tests: Muse Code fixing the open-source Mastodon client Ice Cubes; DaVinci Resolve footage review and cropping.

The 512 GB Mac Studio config was not on sale at launch (planned later in October 2026); the 256 GB model was already priced above $10,000. Hardware math: two M5 Ultra Mac Studios (1 TB) fit the NVFP4 checkpoint; the BF16 checkpoint needs about 1.8 TB (four Mac Studios or equivalent GPUs).

## Benchmarks (Meta-reported, from launch materials)

- JobBench (long-running agent tasks): Muse Spark Open 61.8, Inkling 31.1, Nemotron 3 Ultra 27.2
- Terminal-Bench 2.1: Kimi K3 85.0, GLM 5.3 and Qwen3.8 above 76.4 (exact scores not in coverage), Muse Spark Open 76.4
- Independent runs had not landed by 2026-10-16.

## Access

- Weights: Hugging Face (meta-models org), Amazon SageMaker JumpStart, Google Cloud Model Garden, Microsoft Foundry; community GGUF via Unsloth
- Hosted API: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI
- License: free to download; Muse Model License has limited commercial terms — read before commercial use