Intelligence · agent infrastructure
Meta releases the weights for Muse Spark Open
At a glance
- Meta released Muse Spark Open's model weights on October 15, making the model available to download and adapt.
- Meta reports 816 billion total parameters, about 42 billion active per token and a one-million-token context window.
- The model is also offered through hosted inference providers; a full local setup requires substantial computing resources.
What changed
Meta released Muse Spark Open on October 15, publishing the model weights for an AI system built for coding, tool use and longer agent workflows. Meta reports 816 billion total parameters, about 42 billion active per token, and a context window of up to one million tokens. The weights are available from model distributors, and several companies offer hosted API access. Meta distributes the model under its Muse Model License, which has commercial conditions.
Full read
Meta has made a substantial AI model available outside its own products.
Muse Spark Open, released October 15, is an open-weight version of Meta's Muse Spark system. Developers can download trained model parameters and adapt them for their own projects, or use the model through services that host it. That is a different choice from relying only on a closed AI service that never distributes its weights.
Meta describes Muse Spark Open as designed for coding and longer agent workflows: work that may involve reading a large software project, planning changes, calling tools and checking the results. Meta lists 816 billion parameters in total, about 42 billion active for a given token, with text, image, video and audio inputs and a one-million-token context window. These are Meta's published specifications, not measurements made by AICV.news.
The word open needs qualification. Meta released the weights under its own Muse Model License with conditions on commercial use. The model is not simply unrestricted open-source software. Developers should read the actual license before building a paid service around it.
Nor does a downloadable model mean it runs on an ordinary laptop. Meta says it has optimized Muse Spark Open to run across two Apple Mac Studio computers with M5 Ultra chips; it also lists hosted API providers including Together AI, Fireworks AI, Baseten and others. A small founder can test a hosted route without buying that hardware, but usage prices, performance and fit are provider-specific.
The significance is a new option. Developers can compare a powerful model whose weights are available with closed services, weighing control and customization against license terms, compute requirements and operational work. Independent assessment of Meta's performance claims remains important.
Why it matters here
Our read: Inland Empire and Coachella Valley startups experimenting with custom agents now have another model option they can inspect or adapt instead of depending entirely on a proprietary API. Most small teams will find hosted access more plausible than buying multiple high-end workstations. The release does not by itself establish better costs or results for local businesses.
One question worth testing
For an Inland Empire founder building an agent, does a hosted open-weight model perform the needed task well enough to justify its cost and licensing obligations?
Sources
- Introducing Muse Spark Open, the most capable US open-weight model to date — Meta AI Research (2026-10-15)
- Build with Muse Spark Open — Meta for Developers (2026-10-15)
Not confirmed by us
- Model specifications and performance descriptions come from Meta, not independent AICV.news testing.
- License restrictions and hosted inference prices may vary by use case and provider; the desk has not evaluated deployment economics.
How this was made: researched and written at the desk with AI; every figure checked against its source by agent:desk / GPT-5.6 Sol, with a planted-error test the check caught; published by Sat Singh on October 17, 2026.