Technology · News
Meta Releases Open-Weight AI Model Muse Spark for Coding and AI Agents
Published:
In brief
Meta released Muse Spark Open on October 15, 2026, an open-weight AI model with 816 billion total parameters that the company calls "the most capable US open-weight model to date." Built for coding and agentic workflows, it is also designed to run locally on linked Apple M5 Ultra Mac Studio hardware.
Meta Platforms published the open-weight version of its Muse Spark artificial intelligence model on October 15, 2026, according to a post on Meta AI Research's blog. The company presents it as the most capable open-weight model built in the United States so far, designed specifically for coding tasks and AI agent workflows.
Open-weight means the trained parameters of the model are published for developers to download, inspect and modify, unlike fully proprietary systems whose internal weights stay under the creator's control, according to Reuters reporting carried by Investing.com.
What Muse Spark Open Can Do
Meta's research blog lists 816 billion total parameters with 42 billion active at any one time, and a context window of up to 1 million tokens. The model is described as natively multimodal, handling text, image, video and audio inputs. Meta says the model is designed for extended, multi-step tasks, including complex software-engineering projects that require sustained reasoning.
TechCrunch reported that Meta demonstrated the model fixing bugs in a Mastodon client and performing video-editing tasks in the software DaVinci Resolve, including reviewing and cropping footage. Meta also pairs the model with Muse Code, described by TechCrunch as the company's AI coding terminal, for agentic applications.
Running It on a Mac
According to TechCrunch, running Muse Spark Open locally requires at least 1 terabyte of RAM. Meta's own blog post specifies that reaching this locally means linking two Apple Mac Studio computers equipped with M5 Ultra chips, each supplying up to 512GB of memory, for a combined memory bandwidth of 1.2 terabytes per second. Meta lists compatibility with several inference engines, including SGLang and vLLM among others.
Licensing and Distribution
Meta's research blog states the model is released under the "Muse Model License," which it describes as carrying limited commercial terms while allowing free use for many developers. Model weights are available for download through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted inference access is also offered through partners including Together AI, Fireworks, Baseten, Modal, Groq, Nebius and OpenRouter, according to Meta's blog post.
Background
Investing.com reported that Meta previously released a smaller open-weight model, Muse Glimmer, in August 2026. The outlet also noted that Meta chief executive Mark Zuckerberg has repeatedly argued that advanced AI capability should not be concentrated among a small number of companies, framing the open-weight strategy as a counterpoint to competitors that keep their most capable models closed.
Neither Meta's blog post nor the independent reports reviewed provide third-party benchmark comparisons against specific competing open-weight models; Meta's claim of leading performance is based on the company's own evaluation results.
- Release date: October 15, 2026
- Parameters: 816 billion total, 42 billion active
- Context window: up to 1 million tokens
- Minimum local setup: about 1 terabyte of RAM, reached by linking two Apple M5 Ultra Mac Studios
- License: Muse Model License with limited commercial terms
Questions
- What is Muse Spark Open?
- It is an open-weight AI model released by Meta on October 15, 2026, with 816 billion total parameters, built mainly for coding and AI agent tasks, according to Meta's research blog.
- What does 'open-weight' mean?
- It means the model's trained parameters are published for developers to download, inspect and modify, as opposed to proprietary models whose internal weights remain fully controlled by the creator, per Reuters/Investing.com.
- Can I run Muse Spark Open on my own computer?
- Running it locally requires roughly 1 terabyte of RAM; Meta's blog specifies this means linking two Apple Mac Studio computers with M5 Ultra chips, according to Meta's own blog post and TechCrunch reporting.
- Is Muse Spark Open free to use?
- Meta's blog states it is released under the Muse Model License, which includes limited commercial terms but allows free use for many developers; exact commercial restrictions were not detailed beyond that description.
- How does it compare to other AI models?
- Meta describes it as the most capable open-weight model built in the US to date, based on its own internal evaluations; TechCrunch noted no independent third-party benchmark comparison was provided.
Sources
This piece was prepared with artificial intelligence and passed an automated source check before publication. If you see an error, ask for a correction. Request a correction or reply · How we make this
Related
TechnologyArticle
Tesla's New Roadster Uses SpaceX-Style Thrusters: How the System Works
Tesla unveiled a redesigned second-generation Roadster on October 15, 2026, at a SpaceX test site in Texas, fitted with cold-gas thrusters adapted from rocket hardware. The system stores compressed air in carbon-fibre tanks rather than rocket propellant, and Tesla says it will let the car perform rapid acceleration bursts and aerial maneuvers; production is targeted for 2028.
TechnologyArticle
Tesla's Grünheide Plant: The Plan to Reach One Million Cars a Year
Tesla wants to double the long-term capacity of its only European car plant, in Grünheide, Germany, from a planned 500,000 vehicles a year to one million. The company plans to file permit applications by the end of 2026 for a second production hall and expanded battery output, with construction starting in 2027 at the earliest.
TechnologyNews
Google Cuts Quantum Error Rates 40-Fold With 'Magic State Cultivation'
Google's Quantum AI team has demonstrated a technique called magic state cultivation on its Willow processor, cutting errors in key quantum states roughly 40-fold and reaching 99.99% fidelity. The results, published in the journal Science on October 15, 2026, tackle a long-standing resource bottleneck in quantum computing, but Google says the method still needs to be scaled up before it can support real-world applications.
TechnologyArticle
How AI Chips Like Nvidia's GPUs Are Designed and Manufactured
AI chips are first designed by companies such as Nvidia using specialised software to lay out circuits, then physically built by contract manufacturers such as TSMC through hundreds of lithography, etching and packaging steps that together take more than a year. Equipment maker ASML says demand for the lithography machines this process depends on will outstrip supply at least through 2028, even as Nvidia's own shares fell this week on unrelated, inflation-driven market moves.