Meta Releases Muse Spark Open, an 816B Open-Weight Model Built for Agentic Coding

ByAI Tool Insights Team Hours Categories Generative AI
Muse Spark Open abstract illustration

Meta has released the weights for Muse Spark Open, an 816-billion-parameter open-weight AI model built for coding and agentic work. The company announced the release on October 15, 2026, describing it as the most capable open-weight model to come out of the United States to date. The release came three days after the launch of Muse Wonder, Meta’s closed frontier model, on October 12.

Key specifications

Muse Spark Open is a mixture-of-experts model with 816 billion total parameters, of which about 42 billion are active per token. According to Meta’s announcement, the model uses 256 experts per layer with eight consulted for each token. It offers a context window of 1,048,576 tokens, roughly one million, which Meta frames as large enough to hold an entire software project, including dependencies, legacy code, and thousands of files, in context at once.

The model is natively multimodal, accepting text, images, audio, and video as input and returning text, which suits debugging workflows that mix screenshots, diagrams, error logs, and screen recordings.

Two checkpoint formats are available: the full-precision BF16 checkpoint at about 1.6 terabytes and the compressed NVFP4 version at about 537 gigabytes, the latter reducing memory footprint and supporting larger key-value caches for longer concurrent sessions.

Background and context

Muse Spark Open is the open-weight sibling of Muse Spark, the model family behind Meta AI. Its release continues a busy month for open weights: Reflection AI’s Beam, Mistral’s Large 4 preview, and NVIDIA’s Nemotron 3.5 Super VL have all appeared within recent weeks, and Meta is explicitly positioning this launch as the US answer to high-performing open models from Chinese labs. The company’s own model card compares Muse Spark Open against GLM 5.3 and Kimi K3.

The model is released under Meta’s custom Muse Model License rather than a standard open-source license. The license permits use, modification, fine-tuning, and distribution, and requires products built on the model to display “Built with Muse Spark Open.” Model-as-a-service providers whose aggregate gross revenue exceeds 20 million US dollars in any 12-month period need a separate agreement with Meta. Internal use and research are excluded from that condition.

Running it and using it

Meta’s headline demonstration is running the model on desktop hardware. The company worked with Apple to optimize Muse Spark Open for the MLX framework, and the reference setup is two paired 512GB M5 Ultra Mac Studios. Apple’s unified memory architecture pools the machines into what Meta describes as one system with a terabyte of unified memory and 1.2 terabytes per second of memory bandwidth, though the 512GB configuration was not yet shipping at announcement time.

For most users, the practical route is a hosted API. At launch, hosted API partners included Together AI, Baseten, Fireworks AI, Groq, Modal, Nebius, and OpenRouter, with Groq and OpenRouter named among the options for trying the model without downloading anything. Together AI lists serverless rates of $1.25 per million input tokens and $4.25 per million output tokens. The weights themselves are available for download on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.

NVIDIA published a technical blog on the same day noting the model was trained on GB300 NVL72 Blackwell Ultra infrastructure and is tuned for agentic coding workflows.

Performance and what Meta tested

Meta’s announcement says Muse Spark Open is built for coding and agentic work and excels at long-horizon tasks, complex software engineering, and multi-step reasoning. In its own model card, Meta scores the model at 39 on the Artificial Analysis Intelligence Index, below GLM 5.3 at 45 and Kimi K3 at 44, while on MCP-Atlas it scores 89.2 against 84.2 and 82.3 for the same two rivals. These are vendor-reported comparisons.

For a practical demonstration, Meta tested the setup with Muse Code on real projects: the company says the model fixed bugs and handled full builds for Ice Cubes, an open-source Mastodon client, and processed a DaVinci Resolve editing session from start to finish.

Why it matters

Three things make this release significant. First, it is one of the largest open-weight releases from a US company this year, positioned as the country’s answer to high-performing open models from Chinese labs. Second, the agentic-coding focus matches where the industry is heading: models that sustain long, multi-step engineering workflows rather than just answering questions well. Third, Meta is showing frontier-adjacent capability on desktop hardware, although the two-machine Mac Studio setup required remains far beyond most home users. The Muse Model License also matters for businesses: it is free for most users, but large-scale API hosts must negotiate with Meta.

Sources

The AI Tool Insights editorial team covers the latest developments in AI tools and the AI industry.

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *