What Meta's Muse Spark Open Means for the Future of Open-Source AI

Meta published the weights for Muse Spark Open last week, and the model matters less than the arrangement it confirms. The biggest labs now keep their best model closed and release an open one a tier behind. That sounds like a retreat for open-source AI. We think it's closer to the opposite, because a tier behind is already more than most businesses will ever need.

What Meta actually released

Muse Spark Open went public on 15 October 2026. It's a mixture-of-experts model (one large model split into specialist sub-networks, with only a few switched on for each word), at 816 billion parameters in total and about 42 billion active per token. It takes text, images, video and audio, and has a context window of a million tokens, according to Orca Router's launch breakdown and Unite.AI's coverage.

Most of the headline benchmarks are Meta's own and haven't been reproduced. The audio and video evaluations, meant to be its selling point, haven't been published. Neither has the full safety report. We'd treat it as promising and unproven, which is how we treat every launch for its first month.

Meta has changed what "open" means

For three years Meta was the loudest big company arguing for open models, with Llama as the proof. In April it launched the original Muse Spark as a closed model with no weights at all, its first since setting up Meta Superintelligence Labs. VentureBeat and The Batch both read it as a turn away from Llama. Meta said existing Llama models would stay available. It made no promise about new ones.

Muse Spark Open is the compromise. On the Artificial Analysis index it scores 39, against 48 for the closed Muse Spark 1.3. The best model stays in-house. The open one ships later and a step down.

The licence tells the same story. Downloading means handing over your legal name, date of birth and organisation. Products have to say "Built with Muse Spark Open", and anyone reselling API access past US$20 million a year needs a separate deal with Meta. Fine-tuning and distillation are allowed. Open weights, then. Not open source in the sense the term used to mean, and we'd rather call it what it is.

Open models are already good enough for everyday work

The part that matters is how close "a tier behind" now is. Open models from Qwen, DeepSeek, GLM, Google's Gemma and OpenAI's gpt-oss sit within a few points of the strong closed models on most public leaderboards. The very top is still closed, and the order shuffles month to month. For drafting, sorting, extracting and summarising, the gap has stopped being the deciding factor.

The small end has moved even faster. Stanford's Intelligence Per Watt study found small local models could correctly answer 88.7% of single-turn chat and reasoning queries, up from 23.2% in 2023. Qwen3.6-35B-A3B, released under Apache 2.0, uses about 3 billion parameters per token and runs on a single 24GB graphics card.

That fits what we already believe. The models that exist today can take most of the repetitive admin out of a small business. The bottleneck is building with them properly, and an open model you can run anywhere removes one more excuse.

What changes when anyone can run the model

A closed model is a service. The provider picks the version, sets the price and can retire it with a few months' notice. An open-weight model is a file. Yours. Once you have it, nobody can take it away or change how it behaves under you.

That shifts a few things for the businesses we build for. Sensitive client data can stay on hardware you control. The cost becomes a monthly server bill instead of a per-token one that climbs with use. A workflow that works today still works in two years, because the model under it hasn't been quietly swapped. It's the same reason we say rent commodities and own the core.

Tooling has caught up. llama.cpp, vLLM and MLX supported Muse Spark Open on its first day, so getting an open model running is now a weekend job rather than a research project.

What it means for AI's energy bill

The International Energy Agency reported in April 2026 that data centre electricity use grew about 17% in 2025, against about 3% for electricity overall, and expects it to double by 2030. AEMO's 2026 Electricity Statement of Opportunities puts Australian data centres at about 5 TWh a year now, roughly 3% of the National Electricity Market, rising to 34 TWh by 2035-36 in its central case.

Open models help here in one specific way. They let you pick the smallest model that does the job. A closed app sends every question to whatever model it defaults to. Compute scales with the parameters used per token, and sorting an enquiry with 3 billion of them instead of 42 billion is a large saving repeated thousands of times a week. Text tasks already sit at the cheap end, as our piece on AI slop and energy shows. A right-sized model makes them cheaper again.

Distillation helps as well. Quietly. Licences that allow it let people train small specialist models from a big one, rather than every company training its own from scratch. That's reasoned rather than measured, and we'd rather say so.

Where open and self-hosted falls short

Here's the awkward bit, which the open-source crowd tends to skip. Self-hosting is not automatically greener. The same Stanford study found cloud accelerators deliver at least 1.4 times more intelligence per watt than local hardware running the same model. A data centre serves thousands of people from one card. The workstation under your desk serves one person, and draws power while it waits.

Cheaper inference also tends to mean more of it. The IEA's numbers show it already. Per-task efficiency keeps improving and total demand keeps climbing.

Then there's who controls the future of open models. If the biggest labs decide what gets released and when, open AI depends on their goodwill and their licences. The healthier sign is that the strongest open models now come from several companies in several countries, so no single one can close the door.

What it means for an Australian business

In the short term, almost nothing. Really. Muse Spark Open's full-precision weights are about 1.5TB, and Meta's own desktop demo needed two top-end Mac Studios. Nobody with 30 staff should buy that hardware to answer the shared inbox.

The longer view is more useful. The AI that gives people their week back, the reading, sorting, matching and drafting, never needed a frontier model. Open models mean that work can run on something small, cheap and yours. For most businesses under 20 staff that still means a small open model through a hosted provider rather than a server in the back room.

If you're weighing a hosted API against running a model yourself, our AI agents work covers how we choose, and what AI automation costs has the running costs. The AI glossary covers the terms used here.

People also ask

What is Muse Spark Open?

Muse Spark Open is an open-weight model Meta released on 15 October 2026. It's a mixture-of-experts model with 816 billion parameters, of which about 42 billion are used for each token. It reads text, images, video and audio and replies in text. It ships under Meta's own Muse Model License rather than an open-source licence.

Is Muse Spark Open open source?

Not in the strict sense. The weights are downloadable, but the training data and code aren't published, downloads are gated behind your name, date of birth and organisation, and the licence carries conditions such as a "Built with Muse Spark Open" label. Open-weight is the accurate term.

Is Meta still releasing Llama models?

Meta has said existing Llama models stay available, but it hasn't committed to new ones. Its April 2026 frontier model, Muse Spark, was closed, and Muse Spark Open is the open release that followed six months later.

Can a small business run Muse Spark Open?

Not sensibly. The full-precision weights are about 1.5TB, and Meta's desktop demonstration used two top-end Mac Studios with up to 512GB of memory each. A model in the 3B to 30B range runs on a single workstation GPU and covers most admin work.