Muse Spark Open vs Muse Spark 1.2: The Open Flagship Isn't the Upgrade

The instinct when a lab open-sources its biggest model is to assume the downloadable version is the better one, and in the case of Muse Spark Open against Muse Spark 1.2 that instinct is wrong on the axis people check first. On Artificial Analysis, the proprietary API model scores an Intelligence Index of 40 and the newly opened 816-billion-parameter model scores 39 — within a point, and in the direction that favours the model you rent. The open one is not the stronger model. It is a different bargain, and the bargain is the whole comparison.

The sentence version

Muse Spark 1.2 buys you a fast, cheap, per-token endpoint you never have to operate and whose modalities are all independently confirmed. Muse Spark Open buys you weights you can hold, fine-tune, audit and run without a network, on hardware you have to buy and at a throughput nobody has published yet. Those are different products that happen to share a family name.

Where they actually stand

Line by line, with the side that wins on each dimension stated first.

• Intelligence — Muse Spark 1.2 at an Artificial Analysis Intelligence Index of 40, Muse Spark Open at 39. Independently measured on both sides; a one-point gap is not a reason to choose either model.

• Throughput — Muse Spark 1.2, and by a distance: Artificial Analysis ranks it first of 185 models on output speed at 361.4 tokens per second. Muse Spark Open's speed has not been measured at all, by anyone independent. Nothing on a desk in October 2026 is going to touch 361 tok/s on an 816B mixture of experts; the question is how far behind it lands, and that number does not exist yet.

• Price — Muse Spark 1.2 at $1.25 per million input tokens and $4.25 per million output, with an 88% cached-input discount and a measured $0.97 per Intelligence Index task. Muse Spark Open has no per-token price; its price is a hardware purchase, and Artificial Analysis has not published a cost figure for it.

• Weights — Muse Spark Open, entirely. Gated Hugging Face repositories under the Muse Model License, in BF16 and NVFP4, roughly 1.5TB before runtime overhead. Muse Spark 1.2's weights are not distributed in any form.

• Local operation — Muse Spark Open, on two M5 Ultra Mac Studios with MLX or on a single four-GPU Blackwell node with TensorRT-LLM. Muse Spark 1.2 runs nowhere but a provider.

• Fine-tuning — Muse Spark Open, with Meta shipping TorchTitan support specifically for it. Muse Spark 1.2 is not tunable on your data.

• Modality — contested, and it resolves in Muse Spark 1.2's favour on the evidence available. Artificial Analysis lists 1.2 as accepting text, images, speech and video, and lists the open model as accepting text and images only, while Meta's own materials claim text, image, video and audio for Muse Spark Open. Independent measurement has confirmed the narrower claim; the vendor has asserted the wider one.

• Reasoning — essentially the same arrangement on both. Neither lets you switch reasoning off. Muse Spark Open exposes low, medium, high and xhigh effort levels mapping to 16, 64, 256 and 512 reasoning tokens, and a reasoning-strength hint in the system prompt.

• Context — a genuine tie at 1,048,576 tokens, with one caveat that separates them in practice. On Muse Spark Open, a full million-token context in BF16 costs roughly 266GB of KV cache on top of the weights. On Muse Spark 1.2, context is somebody else's capacity problem.

• Licence — Muse Spark Open under the Muse Model License with limited commercial terms; Muse Spark 1.2 under ordinary API terms. Neither is Apache 2.0, and the open licence is the one you have to actually read.

Read down that list and the shape of the decision is already visible: the open model wins on the things you do to a model, and the hosted model wins on the things a model does to you.

The one number that decides it for most teams

Muse Spark 1.2's headline specification is 361.4 tokens per second, first of 185 models measured. That is not a marketing figure — it is what makes the model usable inside an agent loop, where a coding agent fires dozens of small round-trips per minute and the wall-clock cost of the model is the wall-clock cost of the task.

Muse Spark Open's throughput is unknown. That is the honest state of the evidence, and it is the single largest hole in any comparison written this week. What can be said with confidence is structural: an 816-billion-parameter model activating 42 billion parameters per token on consumer desktop hardware is bounded by memory bandwidth, not by compute. Two M5 Ultra Mac Studios give you 1.2 TB/s and a terabyte of unified memory, which is enough to hold the model and read each token's slice. It is not enough to make an interactive agent feel like an API. If your workload is interactive, this is not a close call. If your workload is batch — overnight repository sweeps, offline evaluation, large-scale extraction where latency is irrelevant and per-token cost is everything — the calculus inverts completely.

Is the desktop actually cheaper?

This is where the open model earns its place, and also where the naive version of the argument fails. The two-Mac-Studio configuration is a capital purchase plus power, cooling and someone's attention. Against that you set tokens you no longer buy.

The comparison only closes at volume, and the reason is that Muse Spark 1.2's cache discount is already doing a lot of work on the hosted side. At 88% off cached input, the model is cheap precisely for the repeated-context pattern that agentic coding produces — the same repository, the same system prompt, the same tool definitions, read again on every turn. A self-hosted deployment does not get a discount; it gets a fixed cost, and the fixed cost is only amortised by usage that is large, steady and predictable. A team running bursty agent traffic on a two-machine desktop setup has bought a fixed cost to serve a variable load, which is the classic way this arithmetic goes wrong.

The honest framing is that self-hosting Muse Spark Open is a capacity decision, not a cost decision. You buy it to guarantee that the model is available, that it cannot be deprecated under you, that your prompts never leave the building, and that you can change its behaviour by fine-tuning. Those are real and sometimes decisive reasons. "It will be cheaper" is a reason that has to be re-derived for your traffic, and for a lot of traffic it simply is not true.

What each one is genuinely for

Muse Spark 1.2 is the model to build a product on. It is fast, it has a measured cost per task, its modality roster is independently confirmed, it is sold per token with a cache discount that suits agentic patterns, and it is routable today. Choosing it is choosing to spend money instead of engineering time, and at $1.25/$4.25 that is often the correct trade.

Muse Spark Open is the model to build a capability on. Fine-tuning is the reason it exists: if the task is narrow, high-volume and well-specified — a codebase-specific agent, a document pipeline for a fixed corpus, a classifier that has to live on-premises — then a tuned 42B-active model you own will beat a general-purpose endpoint you rent, and no amount of prompt engineering closes that gap. The desktop story is the same argument in miniature: the point is not that it runs on a Mac, it is that it runs where you are.

The modality question, and why to care

Meta says Muse Spark Open takes text, images, video and audio. Artificial Analysis, which measured the model, lists text and images. That gap is not pedantry — a natively multimodal open-weight model is a fundamentally different asset from a text-and-image one, and the difference determines whether the model is a candidate for audio and video pipelines at all.

The likeliest explanations are mundane: artificial analysis may have evaluated through an endpoint or configuration that did not expose the extra modalities, or the architecture may accept inputs that the shipped deployment did not exercise. What matters for a decision is that the wide claim is vendor-asserted and the narrow one is measured. If audio or video is the reason you were interested, verify it on your own stack before you commit hardware to it. If it is not, the question is academic — and note that Muse Spark 1.2's speech and video support is confirmed on the independent side, which is not a position the open model can currently claim.

Running them side by side

Most teams that end up with both do not choose between them at all. The working pattern is to keep the hosted model on the interactive path and the open model on the batch and on-premises paths, and to enforce the split in configuration rather than in documentation.

That is what a routing layer is for. OrcaRouter routes Muse Spark 1.2 at Meta's list price with 0% markup, so the hosted side of that split costs exactly what it costs on Meta's own API — a vendor price cut is live on the same key the same day — and it puts the model behind the same interface as 200+ others, with automatic failover when a provider degrades and a routing DSL for moving traffic between models and providers without a rewrite. OrcaRouter does not host Muse Spark Open and does not claim to; a model whose value is that you can download it is a model you run yourself. What the router gives you is the other half of the split, so the interactive path stays on a fast endpoint while the batch path runs on your own hardware.

The verdict

Muse Spark Open vs Muse Spark 1.2 is not a contest between a better and a worse model. It is a contest between owning a model and renting one, and the two sides of that contest are not measured in the same units — one in Intelligence Index points and tokens per second, the other in capital, control and the right to change the weights.

Pick Muse Spark 1.2 if you want the strongest measured capability per dollar of engineering effort, if latency matters, if you want confirmed audio and video, or if you simply need a Muse model on one key today. Pick Muse Spark Open if you need to fine-tune, if the data cannot leave your network, if the workload is large and predictable enough to amortise hardware, or if the availability guarantee is worth more to you than the last point of index score.

The only advice worth giving is to resist the framing that open weights are automatically the more capable option. On the evidence, they are not. They are the more yours.

Muse Spark Open is downloadable now under the Muse Model License; Muse Spark 1.2 is available per token through Meta's API and on OrcaRouter at list price.