The Short Answer
Muse Spark Open is open-weight, not open source. Meta published the weights on 15 October 2026 under its own Muse Model License, which requires attribution and a separate agreement once an AI-model product passes twenty million dollars in revenue. Serving it is a hardware question; owning it is a platform question. On ibl.ai you own all the code and the data and switch models at will.
Meta released Muse Spark Open on 15 October 2026: a natively multimodal sparse mixture-of-experts model with approximately 816B total parameters and approximately 42B active per token.
Each MoE layer holds 256 experts and the router selects 8 per token. Context runs up to 1M tokens. Text, images, audio and video go in; text comes out.
That is a genuinely significant release, and the coverage of it has made two errors worth correcting before anyone writes it into a procurement plan.
Is Muse Spark Open actually open source?
No. Muse Spark Open is open-weight, which is a different thing, and the difference is written into the licence rather than hidden behind it.
The model card states that the weights are governed by the Muse Model License — Meta's own instrument, not Apache 2.0 or MIT. Three provisions change how an enterprise should read it.
Attribution is mandatory. Licensees must prominently display "Built with Muse Spark Open" in user interfaces, documentation, or settings menus. That is a product requirement, not a footnote.
There is a revenue trigger. If the licensee or any affiliate provides a third party with access to AI models as a product or service and aggregate gross revenue "exceeds twenty (20) million US dollars" in any consecutive twelve-month period, the licensee must notify Meta and enter a separate written agreement.
Use is bounded by a usage policy, which prohibits illegal activity, harassment, discrimination, unauthorised professional practice, privacy violations, malware creation, and military or weapons applications. The weights are also not intended for anyone under 18.
None of that makes the release less useful.
It makes it a conditional grant rather than an irrevocable one. An open-source licence does not have a revenue threshold at which you renegotiate with the licensor, and a plan that treats this as permanent is a plan with an unpriced renegotiation in it.
What hardware does it actually take to self-host 816B parameters?
More than the phrase "self-host on your own infrastructure" implies, which is the second correction.
Sparse MoE saves compute, not memory. Only 8 of 256 experts fire per token, so each token costs roughly 42B parameters of arithmetic. But all 256 experts have to be resident, so the memory bill is an 816B model's bill at a 42B model's speed.
The published figures:
| Checkpoint | Size / precision | Reported serving footprint |
|---|---|---|
| Muse-Spark-Open-816B-A42B | ~1.5 TB, BF16 safetensors | 4 nodes of 8 H200 GPUs, tensor parallel |
| Muse-Spark-Open-816B-A42B-NVFP4 | 4-bit, for NVIDIA Blackwell | A single B200 node |
A full 1M-token context adds roughly 266 GB for the KV cache alone at BF16, before weights. NVIDIA has published its own agentic-coding deployment guide for GB300 NVL72, and the vLLM recipe covers the serving path.
So "no tokens to a third-party cloud" is available to organisations that already have a GPU estate, and is a rental decision for everyone else.
The NVFP4 checkpoint is the one that makes this a realistic on-premise conversation, and it is the one to evaluate first. Our self-hosted serving stack guide covers the runtime choices underneath.
How good is it, by its own published numbers?
The model card reports the following, and these are the figures to quote rather than a superlative:
| Benchmark | Score |
|---|---|
| GPQA Diamond | 89.0 |
| Terminal-Bench 2.1 | 76.4 |
| Toolathlon Verified | 72.8 |
| HLE (with tools) | 61.0 |
| Video MME v2 | 61.5 |
| Deep SWE 1.1 | 60 |
Terminal-Bench and Toolathlon are the two that matter for the use case Meta is aiming at, which is long-horizon agentic coding rather than chat. A 1M-token window and 76.4 on Terminal-Bench 2.1 describe a model built to stay coherent across a long tool-using session.
This is also the second Muse Spark story of 2026.
The April release introduced Meta's parallel reasoning architecture and went to select partners under a private API preview, with no downloadable weights. This October release is the open-weight entry in the same line, six months later.
Why does owning the AI stack matter more than which open-weight model wins?
ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.
Every open-weight release moves the lock-in question up one layer.
The model stops being the constraint; the platform holding the model becomes it.
A licence with a twenty-million-dollar renegotiation trigger is exactly the kind of term you want to be able to route around in an afternoon, and that is a platform property, not a model property.
- You own all the code and the data. Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours.
- Model-agnostic. Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform.
- No per-seat pricing. Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ.
- Deploy anywhere. Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.
What should you do about Muse Spark Open this quarter?
Three concrete steps, in order.
Read the licence before the benchmark table. The attribution requirement and the twenty-million-dollar threshold are the two provisions that touch product and finance, and neither appears in any coverage of the parameter count. If you resell AI capability, that clause is the one your counsel needs.
Evaluate NVFP4 first, BF16 second. The 4-bit Blackwell checkpoint on a single B200 node is a procurement conversation. A 1.5 TB BF16 deployment across four 8×H200 nodes is a programme. Starting with the second one is how a pilot becomes a capital request.
Hold the model loosely. Six months separated the closed April release from this open-weight one, and a perpetual-licence platform means the next six-month gap costs you a configuration change rather than a migration. The point of model-agnostic architecture is that you never have to be right about which model wins — see why vendor lock-in is the real risk.
Open weights are good news. They are not the same news as open source, and the distinction is in the licence file both camps keep skipping.
Related: Agents Can Now Reverse Engineer Closed-Source Software — the same "open" ambiguity on the software side, and what a 298,000-star MIT tool does to a vendor's opacity.
Sources: the release date, parameter counts, expert counts, context length, modalities, checkpoint names and serving footprints from Hugging Face's Muse Spark Open announcement, published 15 October 2026; the Muse Model License terms (attribution requirement, the twenty-million-dollar revenue threshold, usage-policy prohibitions, age restriction) and every benchmark score from the Muse-Spark-Open-816B-A42B model card; serving guidance from the vLLM recipe and NVIDIA's GB300 NVL72 deployment post. One claim circulating in secondary coverage was dropped: that this is the strongest US open-weight model ever released. No primary source establishes that ranking, and the model card's own benchmark table is the citable fact instead.