Muse Spark Open pricing: five hosts, one price, and it is closed Muse Spark 1.3's card for a model that scores six points lower

• Every host charges $1.25 in, $0.15 cached, $4.25 out (Together has a free promo endpoint until October 18). That's the closed Muse Spark 1.3's card, and 1.3 scores 45 on Artificial Analysis to Open's 39.
• GLM-5.3 costs about 11% more on our RAG workload and scores six points higher. Renting Open by the token is hard to justify.
• Self-hosting on four B300s gets you to roughly a third of the hosted price, once the box is more than about 36% busy.

Five hosts, one price

Meta put the weights on Hugging Face on October 15 as meta-models/Muse-Spark-Open-816B-A42B, in BF16 and an NVFP4 build. It's a mixture of experts: 816 billion parameters, about 42 billion active per token, 1,048,576 tokens of context, and it takes text, images, video and audio. Meta's own Model API doesn't serve it. The model catalog tells you to self-host or go to a third party. Here's who picked it up:

Per million tokens. Host pricing pages and OpenRouter's endpoint list for meta/muse-spark-open, read October 17, 2026. Speeds are Artificial Analysis measurements.

Every one of them landed on $1.25, $0.15 and $4.25, which is the Standard card Meta charges for the closed Muse Spark 1.3 to the cent. That isn't a coincidence we can prove, but it makes sense: nobody wants to undercut Meta on day two and nobody can charge more for the weaker model. The competition between hosts, for now, is on speed. Baseten serves it at almost twice Together's rate, and Fireworks at about 80 tokens a second. Artificial Analysis also clocks a sixth host, Makora, at 286, but we found no per-token price for it.

One exception, briefly. Together runs a separate promo endpoint, Muse-Spark-Open-816B-A42B-promo, at $0 in and $0 out until October 18. If you want to benchmark Open on your own prompts, today is the day to do it for free.

We didn't find a per-token listing on Bedrock, Vertex or Azure. AWS has it in SageMaker JumpStart, billed per instance-hour, and Google and Microsoft have the weights in their model catalogs for you to deploy.

Same card as the closed model, six points down

PyTorch's launch post says Muse Spark Open shares Muse Spark 1.3's base model, with extra safety post-training so it can run without Meta's serving stack around it. That training cost something. Artificial Analysis scores the open model 39 at xhigh effort. The closed 1.3 at xhigh scores 45.

The open model is cheaper per task, though: $0.84 against $1.37 on AA's run, because it wrote about 110 million output tokens to 1.3's 140 million. Divide one by the other and you're paying 2.2 cents per index point on Open and 3.0 on 1.3. If your task is easy enough that 39 clears it, Open costs less. If you need the extra six points, you can't buy them from Open at any effort setting.

And then there's Meta's Contributor tier, which nobody mentions next to the open release. If you let Meta train on your prompts, it sells 1.3 itself at $0.10 in and $0.20 out, capped at 150 requests a minute. On the RAG month below that's $13.20 against $175.50, roughly 7.5 cents on the dollar, for the higher-scoring model. For a lot of hobby and eval traffic, that's the honest answer to "what's the cheapest Muse Spark".

Against the rest of the open-weight shelf

Say you run a RAG pipeline that stuffs retrieved chunks into every prompt: 120 million tokens in a month, 6 million out, nothing cached because the chunks change per query. Here's that month on each card, with the Artificial Analysis score next to it:

Our arithmetic from each card. AA scores are each model's top published setting on Index v4.3.2 (xhigh for the Muse models, max for GLM-5.3, Kimi K3 and DeepSeek V4-Pro). Different tokenizers count the same text differently, so treat cross-vendor rows as approximate.

The row that matters is GLM-5.3. It costs $194.40 against Open's $175.50, an extra $18.90, and scores 45 to 39. It also beats Open on every coding row in Meta's own launch table: 83.9 to 76.4 on Terminal-Bench 2.1, 67 to 60 on DeepSWE 1.1. Meta printed that comparison itself.

Where Open does fine is against Mistral Large 4 once Mistral's launch discount ends: a point higher for about $13 less a month. Kimi K3 scores five points more and costs $450, about 2.6x as much. Open also takes video and audio input. If your pipeline needs that, check the alternative supports it before you pick it on price.

When running it yourself wins

This is the part the weights are actually for. Meta's card asks for about 700GB of GPU memory in NVFP4 and 2TB in BF16. The smallest NVIDIA setup Meta names is four Blackwell Ultra GPUs; the NVIDIA Dynamo recipe on four GB300s gets about 159 output tokens a second per GPU on an agent trace of 64K tokens in, 400 out, with 90% of the input reused from cache.

Price that request on a host first. 57,600 cached tokens at $0.15, 6,400 fresh at $1.25 and 400 out at $4.25 comes to $0.0183.

Our estimate. B300 rental at Nebius's $9.50 per GPU-hour list price, standing in for GB300; neocloud rates in October mostly sat between $8 and $9, which moves break-even to about a third busy. Throughput from NVIDIA's Dynamo recipe at concurrency 16. Excludes egress, storage and the engineer who keeps it running.

So a box kept busy produces this workload at about a third of the hosted price, and a box idle two-thirds of the time costs roughly what Together would have charged. Below about 1.5 million requests a month, rent the tokens.

Two caveats on our math. The Dynamo numbers were measured at a concurrency of 16; batch harder and the per-GPU rate should go up, which helps self-hosting. And four GPUs won't hold the full 1M context. Hugging Face recommends 8x H200 for that, and the KV cache alone is reported at around 266GB.

What the license lets a host do

The Muse Model License has no monthly-user cap, unlike the old Llama licenses. The gate is on selling the model as a service: a company with more than $20 million in gross revenue over any 12 months has to notify Meta and sign a separate agreement first. Internal use and building it into your own product are exempt.

That clause may be why the host prices look the way they do. Most of the inference companies above are probably past $20 million in revenue, which would mean each has a deal with Meta we can't read. We don't know if price is in it. Everyone also has to show "Built with Muse Spark Open", and the definition of a derivative reportedly reaches models improved with its outputs, so read it before you distill from it.

Our take

For API buyers this release changes nothing yet. If you want a Muse Spark by the token, the closed 1.3 on Meta's API scores higher at the same price, and the Contributor tier is far cheaper if your data can go to Meta. If you want the best open-weight model near $1.25, GLM-5.3 is ahead.

Open makes sense in two cases: you need audio or video input on weights you control, or you have steady volume past a couple of thousand agent calls an hour and a team that can run a Blackwell node. We'd also watch the hosts. Fireworks' preview ends October 29, and the first one to cut below $1.25 will tell you whether that number was a market price or an agreement.

The Muse Spark Open model page will track it if they do. To price your own mix, try the calculator.

Sources

• Hugging Face: Muse-Spark-Open-816B-A42B model card - architecture, license, hardware minimums, Meta benchmarks
• PyTorch: Muse Spark Open and Meta research blog - release date, relation to Muse Spark 1.3, cloud availability
• Meta: build with Muse Spark Open and Hugging Face launch blog - recommended GPU setups, full-context hardware, KV cache size
• Meta: models and pricing and rate limits - not on the Model API, Muse Spark 1.3 Standard and Contributor rates
• OpenRouter: meta/muse-spark-open, Together AI, Fireworks and Baseten - host prices
• Artificial Analysis: Muse Spark Open - Intelligence Index, cost per task, tokens, provider speeds
• NVIDIA Dynamo: Muse Spark Open recipe - per-GPU throughput on GB300
• VentureBeat: Meta unveils Muse Spark Open - license derivative terms