Affiliate Disclosure: Some links on this page are referral links. If you sign up through them, we may earn a commission at no extra cost to you. Our test results are not influenced by commissions.
Meta just did something it hasn’t done since Llama 4: release a genuinely frontier-class model with downloadable weights. Muse Spark Open — 816 billion parameters, 42 billion active per token, 1M-token context — landed on October 15, 2026, three days after Meta’s closed frontier model Muse Wonder. It’s free to download, carries a limited commercial license, and is already live on day-zero hosted APIs.
The short version: this is the cheapest frontier-class API on the market right now ($1.25 in / $4.25 out per million tokens), the strongest U.S. open-weight release to date by independent benchmarks — and a model you almost certainly cannot run yourself. “Free to download” and “runnable by you” are two very different things here. Everything below is from Meta’s release materials, provider pricing pages, and independent benchmark publishers as of October 18, 2026. We have not run this model hands-on (it needs roughly a terabyte of unified memory), and we won’t pretend otherwise.
What Is Muse Spark Open?
Muse Spark Open is the open-weight sibling of Muse Spark, the model family behind Meta AI. Meta calls it “the most capable U.S. open-weight model to date” — a pointed framing, because for the past 18 months Chinese labs have set the pace in open weights unchallenged. The specs: 816B total parameters in a mixture-of-experts architecture with 42B active per token, 1,048,576 tokens of context, natively multimodal (text, image, audio, and video input), built for agentic coding and long multi-step work.
It arrived three days after Muse Wonder, Meta’s closed frontier model — making Meta the only American company currently fielding both a frontier closed model and a frontier open-weight model. (For the pricing side of that sibling rivalry, see our Gemini 4 Argon vs Muse Wonder pricing breakdown.)
Pricing: The Cheapest Frontier-Class API Right Now
“Pricing” was the breakout search query around this launch, so here’s the table — re-verified today against Together AI’s own pricing page, Baseten, and VentureBeat’s comparison table:
| Provider | Input / 1M | Output / 1M | Cached input / 1M |
|---|---|---|---|
| Together AI (serverless) | $1.25 | $4.25 | $0.15 |
| Baseten | $1.25 | $4.25 | $0.15 |
| OpenRouter | $1.25 | $4.25 | — |
For context, here’s what $5.50 per million tokens (blended) buys you versus the closed competition, per VentureBeat’s comparison table:
| Model | Blended $/1M |
|---|---|
| Muse Spark Open (Together AI) | $5.50 |
| GPT-6 Sol | $12.00 |
| Gemini 4 Argon (post-intro) | $24.00 |
| Claude Opus 5.5 | $24.00 |
| Muse Wonder 1.0 | $24.00 |
Spark Open undercuts its own closed sibling by more than 4x on API price. Other day-zero hosts include Fireworks AI, Groq, Modal, and Nebius. Downloading the weights themselves is free.
How to Access It
Hosted API (the realistic path): Together AI, Baseten, Fireworks AI, Groq, Modal, Nebius, and OpenRouter all serve it from day zero. Together’s docs show the standard OpenAI-compatible call pattern against meta-models/Muse-Spark-Open-816B-A42B.
Download the weights: Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Checkpoints ship as BF16 full precision (~1.6TB) and NVFP4 compressed (~530GB); community quants from AMD, Intel, and Unsloth go down to ~2-bit, reportedly runnable on 256GB machines.
Meta’s own API also serves it, priced the same as the closed Muse Spark 1.x line.
Benchmarks: What Independent Numbers Say
Attribution matters here, so: Artificial Analysis scores Muse Spark Open at 39 on its Intelligence Index — well above the median of 19 for open-weight models of similar size, and roughly in the tier of GPT-5.6-class, Claude Opus 5, and Grok 4.6 High systems. Median serving speed across providers: 130.7 tokens/second with 0.73s time-to-first-token.
The honest caveat, and it’s a big one for this site’s audience: on pure coding benchmarks, Chinese open models still lead — Kimi K3, GLM 5.3, and Qwen3.8 all outscore Spark Open’s 76.4 on Terminal-Bench 2.1. “Strongest U.S. open model” is not the same as “strongest open model,” and the gap is the actual story in the open-weights race.
The Mac Story: Technically True, Practically Absurd
The launch’s splashiest claim is that Muse Spark Open “runs on Macs.” The fine print: you need roughly 1TB of unified memory — two 512GB M5 Ultra Mac Studios linked over Thunderbolt 5 (RDMA), running an NVFP4 checkpoint via MLX. Apple co-optimized the stack, and Apple’s John Ternus posted about the launch — the Meta+Apple co-marketing is real.
The reality check: the 512GB Mac Studio isn’t shipping yet, and the 256GB model already costs well over $10,000. Four 256GB units is the alternative configuration. So yes, it runs on Macs — Macs costing more than a car, plural. For everyone else, the hosted API at $1.25/$4.25 is the actual product.
The License Catch
It’s free to download with “limited commercial use” — but read the Muse Model License before building on it:
- Products built on it must display “Built with Muse Spark Open.”
- Companies with over $20M in annual revenue selling API access to it need a separate deal with Meta.
- Open weights, not open source — no training data is released. Our standard framing applies: open weights does not equal uncensored, and it says nothing about the model’s refusal behavior, which we have not tested.
FAQ
Is Muse Spark Open free?
The weights are free to download and free for nearly everyone to use commercially, with the attribution and revenue conditions above. API access is pay-per-token, starting at $1.25/$4.25 per million tokens — the cheapest frontier-class API pricing available.
Is Muse Spark Open open source?
No. It’s open weights: you get the trained parameters but not the training data or full training pipeline. The license also carries commercial conditions, so it’s not OSI-style open source.
Muse Spark Open vs Muse Wonder — what’s the difference?
Wonder is Meta’s closed frontier model (API only, $4/$20 per 1M tokens); Spark Open is the downloadable open-weight sibling at a fraction of the API price. Wonder remains Meta’s strongest system; Spark Open is the one you can host yourself.
Can I run Muse Spark Open on my Mac?
Realistically, no — unless you own two 512GB M5 Ultra Mac Studios. The honest options are community quantized versions on smaller machines or, far more practically, the hosted API.
Is it uncensored?
We haven’t tested its refusal behavior, and open weights say nothing about it either way — a model’s willingness to answer depends on its post-training, not its license. Don’t assume “open” means “unfiltered.”
The Bottom Line
Muse Spark Open matters less for its specs than for its direction: Meta is back in the open-weights game, it’s undercutting the entire frontier API market on price, and it’s doing it while Chinese labs still lead on raw coding benchmarks. If you build agents and want a strong model you can host in your own cloud, this is the most interesting release of the month. If you wanted to run 816B parameters on your laptop — the API is $1.25 per million tokens, and that’s the real story.
Pricing and availability verified October 18, 2026 against Together AI, Baseten, and VentureBeat. Providers adjust prices; check the linked pricing pages before committing spend.