Muse Spark Open: Meta's 816B Model You Can Run Yourself

Meta just did something it swore off with its frontier models: it gave one away. On October 15, 2026, Meta released Muse Spark Open — the open-weight version of the Muse Spark model that had been proprietary since April. Eight hundred sixteen billion parameters, a million-token context window, built for coding and long-running agent work — and weights you can download instead of renting by the token. Meta even demonstrated it running on two Apple Mac Studios instead of a GPU cluster. The pitch is the oldest dream in AI economics: frontier-class intelligence at zero marginal cost per task. The reality, as always, has a hardware bill attached. Here is what Meta announced, what it actually costs to run, where it honestly scores, and the cheapest way to touch this generation of Muse for free.

The short answer

Meta open-sourced the weights of a frontier-class model. Muse Spark Open is an 816-billion-parameter mixture-of-experts model (42 billion active per token) with a 1-million-token context window, natively multimodal across text, image, video, and audio, tuned for coding and agentic work. It is the same base model as Muse Spark 1.3 with additional safety post-training, per Meta.

The headline that matters for your wallet: you can run it without paying per token — self-hosted, or through a growing list of API providers (Together AI, Fireworks, Baseten, Modal, Groq, Nebius, OpenRouter, per Meta) competing on price. The headline that matters for your budget: self-hosting needs about a terabyte of memory. Meta's own demo ran on two Mac Studio desktops with M5 Ultra chips, each carrying 512 GB of unified memory, linked through Apple's MLX framework.

So the "$0 per task" dream is real — if you already own the supercomputer. For everyone else, the math is more interesting than the slogan, and it's below.

Sources: Reuters: Meta releases open-weight Muse Spark model for coding, agentic tasks (Oct 15, 2026); TechCrunch: Meta launches an open-weight AI model that runs on Macs (Oct 15, 2026); SiliconANGLE: Meta releases 816B-parameter Muse Spark Open for coding and agent work (Oct 15, 2026); MarkTechPost: Meta Releases Muse Spark Open — specs and deployment details. Primary source: Meta's October 15, 2026 announcement.

What Meta actually announced

Specs via Meta's announcement as reported by Reuters, TechCrunch, and SiliconANGLE. The practical translation: sparse activation means it computes like a 42B model but stores like an 816B one. The weights have to live somewhere, and that somewhere is measured in hundreds of gigabytes.

The $0-per-task dream vs the 1TB reality

This is the part our readers care about, so let's do it in our unit: cost per finished task.

Route 1 — self-hosted: $0 marginal per task, five figures of hardware. Once the weights are loaded, every additional task costs you electricity. But "loading the weights" means ~537 GB (NVFP4) to ~1.65 TB (BF16) of fast memory before KV cache and overhead. Meta's reference setup is two maxed-out Mac Studios acting as one machine. Quantized community builds (e.g., ~300 GB 2-bit via Unsloth, per press reporting) lower the bar — but you are still buying serious hardware to get to zero.

Route 2 — hosted APIs: the price war is the point. With Together, Fireworks, Groq, OpenRouter, and others all serving the same open weights, providers compete on per-token price instead of Meta setting it. Open-weight releases historically push inference prices down fast — check current listings before assuming any figure, because they move weekly.

Route 3 — the Muse app: still the cheapest seat. Here's the honest bottom line for non-developers: Meta monetizes the agent through subscriptions and referral bonuses, not per-token metering. The 1-billion-token sign-up bonus (details at the end) buys more Muse-family intelligence at zero marginal cost than any self-hosting rig most readers will ever buy. The open weights are a developer story; the bonus is a consumer story.

The pattern to remember: open weights move the cost from per-task to upfront. If you run thousands of tasks a month, self-hosting wins. If you run dozens, the API — or the bonus — wins. Do the multiplication for your own volume before romanticizing either side.

How it actually scores

All figures below are vendor or press-reported — Meta's own tables, not independent verification. Treat them as directional.

• JobBench (agentic work): 61.8 in Meta's table — more than double the 31.1 Meta lists for Thinking Machines' Inkling and ahead of NVIDIA's Nemotron 3 Ultra at 27.2 (SiliconANGLE, reporting Meta's materials).
• Terminal-Bench 2.1 (coding): 76.4 — behind Kimi K3 (85.0), GLM 5.3 (83.9), and Qwen3.8 (82.0). The Chinese open models still lead on raw coding, per the same reporting.
• MCP-Atlas: 89.2, the top score in Meta's comparison table (MarkTechPost).

The honest read: best US open-weight model is a defensible claim; best open model, full stop, is not — at least on coding, where three Chinese models outscore it. And "most capable" rests on Meta's own test selection until independent evaluations land. What the scores do establish is the tier: this is genuinely frontier-adjacent, not a distilled toy. An 816B MoE at 42B-active is priced — in compute terms — like a mid-size model while reasoning like a large one. That is the whole economic argument for MoE, and it holds here.

The license catch

Read this before you build a business on it: Muse Spark Open is open-weight, not open-source. It ships under Meta's own Muse Model License, which Meta describes as carrying "limited commercial terms" (TechCrunch, TechBooky). That means: many developers can use it free, but the license is not the MIT/Apache-style grant the phrase "open source" implies. If you plan commercial deployment, read the actual license terms — "downloadable weights" and "do whatever you want" are different things, and Meta chose the former deliberately.

This is also the quiet end of an era: the company that built Llama's open-source goodwill now ships its best models under custom licenses with commercial limits. The weights are open; the business model isn't.

What to do this week

• Developers: try it hosted before you buy hardware. Spin it up on one of the listed providers (Together, Fireworks, Groq, OpenRouter) against your current model on 30–50 of your own tasks. Compare success rate and cost per finished task — not parameter counts.
• Don't buy two Mac Studios on launch day. Watch the quantized community builds mature for a month; the 300 GB-class setups are where this gets interesting for mortals.
• Read the license if money is involved. "Limited commercial terms" needs your lawyer's eyes, not a blog's summary.
• Everyone else: feel the generation through the bonus. The Muse referral bonus — 1 billion free tokens for joining with a code — remains the cheapest way to experience what this model family does. The 48-hour redemption clock is real; details below.
• Track our ranking. When Artificial Analysis scores Spark Open independently, we'll place it on the cost-per-task ranking — that will show whether open weights actually break the price curve or just move it.

Frequently asked questions

What is Muse Spark Open?

Muse Spark Open is the open-weight version of Meta's Muse Spark AI model, released October 15, 2026. It has 816 billion total parameters (42 billion active per token via mixture-of-experts), a 1-million-token context window, and is natively multimodal across text, image, video, and audio. It is built for coding and long-horizon agentic work, based on Muse Spark 1.2 with additional safety post-training.

Is Muse Spark Open free?

The weights are free to download, but running it is not free: self-hosting needs roughly 537 GB (NVFP4) to 1.65 TB (BF16) of memory — Meta demonstrated it on two 512 GB Mac Studios. Hosted API providers (Together AI, Fireworks, Groq, OpenRouter and others) charge per token at competitive rates. The cheapest way for most people to try the Muse model family remains the 1-billion-token referral bonus in the Muse app.

Can I run Muse Spark Open on my Mac?

Only on very high-end hardware. Meta's reference setup is two Mac Studio desktops with M5 Ultra chips (512 GB unified memory each) linked via Apple's MLX framework — about a terabyte of memory total. Community quantized builds (~300 GB) lower the bar but still need serious hardware. A regular laptop or desktop cannot run the full model.

How does Muse Spark Open score on benchmarks?

Per Meta's own tables (vendor data): 61.8 on JobBench for agentic work and 89.2 on MCP-Atlas. On the independent-style Terminal-Bench 2.1 coding test it scores 76.4, behind Kimi K3 (85.0), GLM 5.3 (83.9) and Qwen3.8 (82.0). Independent evaluations had not landed at publication time.

Is Muse Spark Open open source?

No — it is open-weight, not open-source. It ships under Meta's Muse Model License with what Meta calls 'limited commercial terms.' Developers should read the license before commercial use; it is not an MIT/Apache-style grant.

Where can I download Muse Spark Open?

Weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models, per Meta. Hosted API access is offered by Together AI, Fireworks AI, Baseten, Modal, Groq, Nebius, and OpenRouter, among others.

Try the Pareto winner free

Muse Spark 1.3 topped our intelligence ranking. Try it with 1 billion free tokens.

Code: S5V9L9

Muse is Meta's AI agent that finishes real errands — email, forms, bookings, research — and keeps working after you close the app. It scored 48 on the Artificial Analysis Intelligence Index, the highest in our 7-model cost-per-task ranking. Join at muse.ai/join, then open Settings > General > Redeem invite code and enter the code above. You get 1 billion free tokens; the person who referred you gets the same.

48-hour deadline: the code must be redeemed within 48 hours of joining, or the tokens are gone for good. Do it the same day.

Claim 1B free tokens with S5V9L9

Full step-by-step: our Muse review.

Want the short version for a specific app? Start with our full directory of 13 reviewed tools and apps or see how we rank them.