Hermes Vector

AI & SITrending

Meta Releases Muse Spark Open, an 816B-Parameter Open-Weight AI Model

Meta has released Muse Spark Open, an 816-billion-parameter open-weight AI model designed for coding and agentic tasks. The model features a 1-million-token context window and is optimized to run on two Apple M5 Ultra Mac Studios.

By Hermes-Vector AI Desk

Also in: Español · Português · Français · Deutsch · हिन्दी

Checked against 6 sourcesresearch.meta.aiThe Guardiantech-insider.orgfinance.yahoo.comMarkTechPostshattered.io
Illustration for: Meta Releases Muse Spark Open, an 816B-Parameter Open-Weight AI Model

In brief

  • Meta released Muse Spark Open, an 816B-parameter open-weight AI model with 42B active parameters per token.
  • The model features a 1-million-token context window and is optimized to run on two Apple M5 Ultra Mac Studios.
  • Muse Spark Open is available via seven hosted inference platforms and four weight-distribution channels.
  • Meta claims it is the most capable US open-weight model, though it trails some Chinese models on specific coding benchmarks.

Meta Releases Muse Spark Open, an 816B-Parameter Open-Weight AI Model

Meta released Muse Spark Open on October 15, 2026, a new open-weight artificial intelligence model the company describes as "the most capable US open-weight model to date." The model is a natively multimodal mixture-of-experts system with 816 billion total parameters and 42 billion active parameters per token. It supports a context window of 1 million tokens and accepts text, images, video, and audio as input, producing text output.

Architecture and Capabilities

According to Meta, Muse Spark Open is built for coding and agentic work, excelling at "long-horizon tasks, complex software engineering, and multi-step reasoning." The model uses a mixture-of-experts architecture with 256 experts per layer, where 8 experts are selected for each token. This design allows the model to maintain high capacity while keeping inference costs lower than dense models of equivalent scale.

The model card, published through Hugging Face, details that Muse Spark Open includes a 1.8-billion-parameter Perception Encoder for visual input and a vocabulary of 202,048 tokens. Training data covers more than 100 languages. Meta states that the open-weight version includes additional safety post-training to address risks associated with public distribution.

Apple Silicon Integration

A key feature of the release is its optimization for Apple hardware. Meta says Muse Spark Open is the first model in its class to run on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework. Apple’s unified memory architecture allows the M5 Ultra's integrated GPU to use up to 512 GB per Mac Studio, enabling two machines to hold the entire model in memory. Meta describes this setup as two Mac Studios behaving like "one supercomputer with a terabyte of unified memory."

Meta demonstrated agentic workflows running entirely on-device, including shipping real iOS changes to the open-source Mastodon client Ice Cubes and driving DaVinci Resolve to review footage frame by frame.

Distribution and Licensing

Muse Spark Open is available through seven hosted inference platforms: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Model weights are available for download on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. According to MarkTechPost, the weights ship in BF16 and NVFP4 formats.

The model is released under the Muse Model License, which includes limited commercial terms. MarkTechPost reports that products using the model must display "Built with Muse Spark Open," and Model-as-a-Service providers with revenue above $20 million in a 12-month period require a separate agreement with Meta. Internal use, research, and powering own product features are exempt from the revenue clause.

Benchmark Performance

Meta compares Muse Spark Open against other leading open-weight models, including Kimi K3, GLM 5.3, and Qwen3.8. According to Meta's evaluation, Muse Spark Open scores 89.2 on MCP-Atlas, ahead of Qwen3.8 (84.5) and GLM 5.3 (84.2). However, on Terminal-Bench 2.1, a coding benchmark, Muse Spark Open scores 76.4, trailing Kimi K3 (85.0), GLM 5.3 (83.9), and Qwen3.8 (82.0).

Why it matters

The release of Muse Spark Open marks a strategic shift for Meta, which has historically focused on open models but recently pushed to compete with closed models from OpenAI and Anthropic. By releasing a highly capable open-weight model, Meta aims to reclaim ground in the open-weight market, which has been dominated by Chinese labs such as Alibaba, Moonshot AI, and Z.ai. The integration with Apple hardware also signals a new partnership between the two tech giants, potentially lowering the barrier for developers to run frontier-scale models locally.


Sources

This article was drafted with AI assistance and checked against the sources above. Company claims are reported as claims. Cover image is AI-generated.

Questions readers ask

What is Muse Spark Open and when was it released?
Muse Spark Open is an 816B-parameter open-weight AI model released by Meta on October 15, 2026. It is described as the most capable US open-weight model to date and supports a context window of 1 million tokens.
What hardware is required to run Muse Spark Open locally on Apple devices?
The model can run on two Apple M5 Ultra Mac Studios using the MLX framework. Apple’s unified memory architecture allows each Mac Studio to use up to 512 GB of memory, enabling the two machines to hold the entire model.
What are the licensing terms for commercial use of Muse Spark Open?
The model is released under the Muse Model License, which requires products to display 'Built with Muse Spark Open.' Model-as-a-Service providers with revenue above $20 million in a 12-month period must sign a separate agreement with Meta.
How does Muse Spark Open perform on benchmarks compared to other models?
Muse Spark Open scores 89.2 on MCP-Atlas, ahead of Qwen3.8 and GLM 5.3. However, it scores 76.4 on Terminal-Bench 2.1, trailing Kimi K3, GLM 5.3, and Qwen3.8.

The day's checked news, by email

One short email a day with the top stories. Free, and you can unsubscribe in one click.

Or follow for the latest news:

Advertisement

How do you feel about this story?

Comments 0

No comments yet. Start the discussion.

AI & SI

Microsoft warns AI agents pose insider threat risks to enterprises

Microsoft introduced new security features to monitor AI agents, warning that autonomous systems can pose insider risks even without being hacked. The company urges separating intelligence supply from control authority.

Checked against 4 sourcesTechCrunchchosun.comdevdiscourse.com+1 more