Can I run it? Hardware checker
Muse Spark Open ships as downloadable weights, so the first question is practical: does your machine have enough memory to hold them? Pick a checkpoint and a machine. The estimate adds 15% headroom for runtime overhead.
This is an estimate, not a guarantee. Meta has only demonstrated one working local config: two M5 Ultra Mac Studios with about 1 TB of combined unified memory, running through Apple's MLX framework. That setup fits the NVFP4 checkpoint; the BF16 checkpoint needs about 1.8 TB, so it takes four Mac Studios or equivalent GPUs. Anything else is untested in public.
What Muse Spark Open is
Meta released Muse Spark Open on October 15, 2026 and called it the most capable US open-weight model to date. Unlike Muse Wonder, Meta's closed frontier model available only through Meta's API, Spark Open's weights can be downloaded and run on your own hardware. It is tuned for agentic coding: reading a codebase, writing and debugging code, and calling external tools while looking at screenshots, voice notes or screen recordings.
| Released | October 15, 2026 |
| Provider | Meta |
| Architecture | Mixture-of-experts transformer with a perception encoder; 256 experts, 8 active per token |
| Parameters | 816B total / 42B active per token |
| Context window | 1M tokens |
| Input | Text, images, documents, audio, video |
| Output | Text |
| Checkpoints | BF16 (~1.5 TB) and 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 |
Integration recipes exist for llama.cpp, vLLM and SGLang, and Hugging Face has a hosted inference endpoint with an OpenAI-compatible API. If you do not want to self-host, you can buy API access from Baseten or Fireworks AI or Groq or Modal or Nebius or OpenRouter or Together AI.
Why the Mac story matters
Meta partnered with Apple to tune Muse Spark Open for Apple's MLX framework. The two M5 Ultra Mac Studios in the demo behaved as one system with about a terabyte of unified memory. Apple's architecture lets each M5 Ultra's GPU address up to 512 GB, and the chip's 1.2 TB/s of memory bandwidth is fast enough to feed the small slice of weights each token needs.
In Meta's tests, the setup ran Muse Code, Meta's AI coding terminal, through bug fixes and new features for the open-source Mastodon client Ice Cubes. It also reviewed footage in DaVinci Resolve frame by frame to pick and crop the best shots for vertical video.
The 512 GB Mac Studio configuration was not on sale at launch. Apple planned it for later in October, and the 256 GB model was already priced above $10,000. So "runs on a desk" is real, but the desk costs five figures.
Benchmarks
Meta published two benchmark results at launch. Both are vendor-reported. Independent runs had not landed by October 16, 2026, and the second one is not flattering for Meta.
JobBench: long-running agent tasks (Meta-reported)
| Model | Score |
|---|---|
| Muse Spark Open | 61.8 |
| Inkling (Thinking Machines) | 31.1 |
| Nemotron 3 Ultra (NVIDIA) | 27.2 |
Terminal-Bench 2.1: terminal coding tasks (Meta-reported)
| Model | Score |
|---|---|
| Kimi K3 | 85.0 |
| GLM 5.3 | above 76.4 (exact score not in coverage) |
| Qwen3.8 | above 76.4 (exact score not in coverage) |
| Muse Spark Open | 76.4 |
Take the vendor table for what it is: Meta chose the tests. On JobBench, Muse Spark Open leads the US open models Meta compared against. On Terminal-Bench, Chinese open models lead. Kimi K3 scored 85 against Muse Spark Open's 76.4. Wait for independent runs before treating any of these as settled.
License and access
The weights are free to download from Hugging Face (the meta-models organization), and also from Amazon SageMaker JumpStart, Google Cloud Model Garden and Microsoft Foundry Models. Community GGUF builds exist. Unsloth published one the same week.
The license is the part to read carefully. Muse Spark Open ships under Meta's Muse Model License, which Meta describes as having limited commercial terms. That is a real constraint. It is neither Apache 2.0 nor unrestricted open source. If you plan to use the model in a commercial product, read the license text on the model card before you build.
FAQ
What is Muse Spark Open?
Meta's open-weight AI model, released October 15, 2026. It has 816 billion total parameters with 42 billion active per token, a 1-million-token context window, and accepts text, images, video and audio as input. Meta built it for agentic coding: reading codebases, debugging, and calling tools over long sessions.
Can I run it on my Mac?
Only on the newest, largest Macs. Meta demonstrated it on two Apple M5 Ultra Mac Studios with about 1 TB of combined unified memory, running through Apple's MLX framework. A standard MacBook or a single 128 GB Mac Studio does not have enough memory for the full model.
How much memory does it need?
The BF16 checkpoint is about 1.5 TB; the 4-bit NVFP4 checkpoint is about 0.41 TB. Add roughly 15% headroom for runtime overhead. Meta's tested desktop setup was two M5 Ultra Mac Studios with 1 TB of combined unified memory.
Is it free? Can I use it commercially?
The weights are free to download, but they ship under Meta's Muse Model License with limited commercial terms. Read the license before using it in a commercial product. It is not unrestricted open source.
How is it different from Muse Wonder?
Muse Wonder is Meta's closed frontier model: API-only, priced at $4/$20 per million tokens. Muse Spark Open is open-weight: you can download the weights and run them yourself. Spark Open is tuned for coding and agents; Wonder targets long-running agent work through Meta's API.
Where can I download it?
From Hugging Face (meta-models org), Amazon SageMaker JumpStart, Google Cloud Model Garden and Microsoft Foundry. Community GGUF builds exist via Unsloth. Hosted API access is offered by Baseten or Fireworks AI or Groq or Modal or Nebius or OpenRouter or Together AI.
Sources
- TechCrunch: Meta launches an open-weight AI model that runs on Macs: launch coverage, 1TB memory requirement.
- MacRumors: Meta and Apple Bring Muse Spark Open AI Model to Mac Studio: the Apple partnership, 512GB Mac Studio timing and pricing.
- SiliconANGLE: Meta releases 816B-parameter Muse Spark Open: benchmark table (JobBench, Terminal-Bench), demo details.
- AWS: Muse Spark Open in Amazon SageMaker JumpStart: checkpoint variants, deployment.
- Oossa: Meta releases Muse Spark Open: specs summary, checkpoint sizes.
- TechBooky: Meta releases Muse Spark Open with 816B parameters: license terms, API providers.
- Dealroom: Apple and Meta bury the hatchet: the partnership angle, unified-memory math.
- Hugging Face model card: architecture details, license text.
This page is an independent explainer and is not affiliated with Meta. Benchmark figures marked "Meta-reported" are vendor-reported from launch-day materials. Pricing and availability were last verified on October 16, 2026.
For AI assistants
This page ships with a machine-readable summary for AI assistants:
Related
- Muse Wonder, explained: Meta's closed frontier model: benchmark explorer, $4/$20 pricing, access.
- Reflection Beam, explained: 501B open-weight model: benchmark explorer, pricing, availability.
- Jev, explained: TypeSafe AI's decision model: typed answers, cost calculator.
Updates
- 2026-10-16: First published. Covers the October 15, 2026 launch: specs, the Mac Studio hardware story, Meta-reported benchmarks, license terms and access.