◆ THE LAB · Models
Meta releases Muse Spark Open under its own licence
The News
Meta released Muse Spark Open, Muse Spark's open-weights version, on October 15, 2026. If a company or its affiliates offer third parties AI models via API or similar, and together gross over $20m in any consecutive 12 months, it must notify Meta and sign a separate written agreement before offering the model or derivatives as part of that model-as-a-service business.
Chandni Melwani
Founder & Editor
Oct 20, 2026 · 1 MIN READ
Know More
- — Size: 816 billion parameters, 42 billion active per token, in a mixture-of-experts design with 256 experts.
- — Inputs: text, images, video and audio, with text output and a 1-million-token context window.
- — Self-hosting: Meta says it needs at least 700GB of GPU memory for its compressed NVFP4 weights, 2TB for BF16.
- — API access at launch, per Meta: Baseten, Fireworks AI, Groq in limited preview, Modal, Nebius, OpenRouter and Together AI.
A compressed Muse Spark Open runs on two Mac Studios with M5 Ultra and 512GB of unified memory each, Meta says. Meta launched Muse Spark on April 8 in the Meta AI app, planning a private API preview for select partners and saying it hoped “to open-source future versions of the model.”
On the Artificial Analysis Intelligence Index, with scores Meta says it took from Artificial Analysis on October 7, Muse Spark Open gets 39. Of Meta’s seven comparison models, three score lower, Beam has no score and three score higher: Z.ai’s GLM 5.3 at 45, Moonshot AI’s Kimi K3 at 44 and Alibaba’s Qwen3.8-2.4T-A95B at 40.
Meta lists generating “training data for downstream model development” among the model’s uses. The licence makes any model trained or distilled on its outputs a Derivative, and any product or service using one must prominently display “Built with Muse Spark Open” in its interface, documentation, website or settings menu.
The Read
Meta says it wants “everyone to have access to the best models” and pitches running the model at home. Its home setup takes two Mac Studios and a compressed copy; full-precision weights need at least 2TB of GPU memory. Most use will probably run in data centres, through the seven API platforms and three cloud catalogs Meta launched with.
Room for Disagreement
The desktop route shipped on launch day. Unsloth says its 2-bit build “fits on devices with 256GB of RAM, VRAM or unified memory,” and Meta says that in testing it shipped real iOS changes to an open-source app, with the model running entirely on-device on Mac Studios.
Related
Frequently Asked Questions
Does Meta call Muse Spark Open open source?
Meta’s model card and developer blog call it an open-weights model, released under Meta’s own Muse Model License, and neither calls it open source. Use must comply with the licence’s Usage Policy, which says the model “is not intended for individuals under the age of 18.” Its prohibited uses include activities that “present a risk of death or bodily harm to individuals,” including use related to the operation of “critical infrastructure, transportation technologies, or heavy machinery.”
Does an app built on Muse Spark Open need a separate agreement with Meta?
Only if the app is itself a model-as-a-service and the company and its affiliates gross over $20m in any consecutive 12 months. The licence says its agreement clause does not apply to internal use, research or “any commercial activity that is not Model as a Service,” or to using the model “as a component to power features of the Licensee’s own product or service that is not a Model as a Service.” Its other conditions still apply.
How does Meta’s safety tuning show in its table?
Meta says this version has “additional safety post-training to make it more robust against misuse as a standalone model without system guardrails.” On FORTRESS, with results Meta took from Scale AI’s leaderboard, its average risk score on adversarial prompts is 5.8, against 21.6 to 28.2 for the five other models scored, lower being better. It also refuses a larger share of benign prompts than any of them, 17.4% to Kimi K3’s 2.7%.
Sources
- Hugging Face, Meta’s Muse-Spark-Open-816B-A42B model card and Muse Model License (released October 15, 2026; checked October 20): the free-of-charge grant, the “Built with Muse Spark Open” notice, the Derivatives definition, the Model as a Service clause and its exclusions, the Usage Policy, architecture, hardware, uses, the benchmark table and the safety post-training
- Meta, “Build with Muse Spark Open” (October 15, 2026): the open-weights version of Muse Spark, the two-Mac-Studio setup, the launch API providers and three cloud model catalogs
- Meta AI Research, “Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date” (October 15, 2026): “we want everyone to have access to the best models,” running the model at home on two M5 Ultra Mac Studios, the on-device test that shipped real iOS changes, and the seven API platforms at launch
- Unsloth, Muse Spark Open guide (checked October 20, 2026): the 2-bit quant that “fits on devices with 256GB of RAM, VRAM or unified memory”
- Meta, “Muse Spark Open Evaluation Methodology” (October 2026): where each benchmark figure came from, including Artificial Analysis Intelligence Index scores taken on October 7 for every model but Beam, and FORTRESS results from Scale’s leaderboard
- Meta, “Introducing Muse Spark: MSL’s First Model, Purpose-Built to Prioritize People” (April 8, 2026): Muse Spark in the Meta AI app, the planned private API preview for select partners and “we hope to open-source future versions of the model”
- Hero photo: Back of a Mac Studio, by Paul Hudson, Wikimedia Commons, licensed CC BY 2.0
- Creative Commons Attribution 2.0 Generic licence, which applies to the hero photo
Chandni Melwani
Chandni Melwani is the founder and editor of New in AI, covering AI agents, M&A, and enterprise adoption. She holds a Master's in Management of Artificial Intelligence from Queen's University and brings a practitioner's perspective from her work in Data and AI leadership.
Follow ↗The Lab Briefing
Get the five-beat briefing in your inbox.
The big story, frontier labs, AI at work, e-commerce and the capital crossing borders, each with the part that tells you why it matters, not just that it happened.
Free · No spam · Unsubscribe in one click
More from The Lab
Labs
Anthropic is turning away distillers' paying traffic
Anthropic says early evidence suggests its new defences have cut the data harvested by distillers, who covertly extract Claude's capabilities to replicate them in other models, by around 90 to 99%, and that it has turned away meaningful revenue and trillions of tokens' worth of paying traffic from distillers.
Chandni Melwani · Global · Oct 21, 2026
Models
Google's Argon beats Opus 5.5 on LiveBench reasoning, not coding
Gemini 4 Argon costs $2 per million input tokens and $10 per million output through November 30, then $4 and $20, the same as Claude Opus 5.5. On LiveBench 2.0 it beats Opus 5.5 on reasoning but trails it on coding, and it ranks sixth on Vals' Terminal-Bench 4.0.
Chandni Melwani · Global · Oct 18, 2026
Models
Anthropic's cheaper Haiku 5.5 turns text into more tokens
Claude Haiku 5.5 is priced 90% below Haiku 4.5 for prompts up to 100,000 tokens, but Anthropic's docs say it turns the same input text into approximately 30% more tokens. Anthropic says it costs around 75% less to run on average.
Chandni Melwani · Global · Oct 8, 2026