Muse Spark Open vs Muse Spark 1.3: Nine Index Points Is a Lot to Give Up for Free

Here is the uncomfortable comparison in Meta's own family. Muse Spark Open, released October 15, 2026, is the 816-billion-parameter open-weight model anyone can download; Independent measurement puts it at an Artificial Analysis Intelligence Index of 39. Muse Spark 1.3, the proprietary API model Meta shipped on September 2, 2026, sits at 48. Nine points is not a rounding difference — it is roughly the distance between one capability tier and the next, and it is the number that makes "just run the open one instead" a much worse piece of advice than it sounds on launch day.

Start with the scenario, not the spec sheet

Three teams looked at the Muse Spark Open release and asked whether they could cancel something.

The first runs a nightly pipeline that reads a fixed corpus and emits structured records. Latency is irrelevant, volume is enormous, the schema never changes. This team should take the open weights and stop paying per token.

The second runs an interactive coding agent that fires a dozen small round-trips per user request. This team should not, and the reason is throughput and reasoning cost, both of which are about to get a lot more concrete.

The third runs hard multi-step reasoning — the kind where the model has to hold a long chain of inference, call tools inside it, and recover from its own dead ends. This team should not either, and for them the nine-point gap is not an abstraction. It is the difference between a workflow that completes and one that has to be babysat.

Everything below is about telling those three apart without guessing.

What the nine points are made of

The index is a composite, and composites hide structure. The structure here is mostly where you would expect it, with one surprise.

• Agentic and tool-use work — Muse Spark 1.3's strongest ground. Meta's own launch figures for 1.3, vendor-reported and measured inside its own harness, put Terminal-Bench 2.1 at 83.9 and DeepSWE 1.1 at 68. Against Muse Spark Open's vendor-reported 76.4 and 60, that is a consistent seven to eight point lead on the evaluations closest to real agent behaviour, and it lines up almost exactly with the independent index gap.

• Closed-book reasoning — the wider gap. Muse Spark Open's own scorecard shows its weakest row is HLE-Diamond without tools, where Meta reports 26.1. Muse Spark 1.3 was built as a reasoning-first model; the "max" reasoning configuration in its name is not decoration.

• Knowledge breadth — narrower than the headline suggests, and Artificial Analysis scores both models on an omniscience axis where the difference is smaller than the index delta implies. This matters because it means the gap is concentrated in reasoning depth and agent endurance, not in how much the model knows.

• Speed — Muse Spark 1.3 measures 127.7 tokens per second, ranked 41st of 185 on Artificial Analysis. Muse Spark Open's speed has not been published by anyone independent. A self-hosted 816B mixture of experts on two desktop machines will be slower than 127.7 tok/s, by how much is unknown, and this is the second-largest hole in the evidence alongside cost.

• Price — Muse Spark 1.3 sells at $1.25 per million input and $4.25 per million output with an 88% cached-input discount, producing a measured $1.60 per Intelligence Index task. Muse Spark Open has no per-token price; its cost is hardware, and no independent cost-per-task figure exists for it.

One line in that list deserves a second look, because it is the argument most often got wrong. Muse Spark 1.3 is a verbose reasoner. Artificial Analysis measured it consuming roughly 170 million tokens across the evaluation, against about 130 million for Muse Spark 1.2 — which is why 1.3 costs $1.60 per task at the same headline token prices that produce $0.97 for 1.2. The frontier reasoning tier charges you twice: once for the capability and once for the tokens it burns thinking its way there.

The reasoning dial is where the money is

Both models are reasoning-first and neither will let you turn it off, but they expose the effort differently, and that difference is the most practically useful thing on this page.

Muse Spark Open publishes its effort ladder in the open: low, medium, high and xhigh map to 16, 64, 256 and 512 reasoning tokens, and a reasoning-strength hint in the system prompt does the same job. That is a dial you can turn from your own code, with a known cost per setting, and because you are paying in electricity and wall-clock rather than in dollars per output token, turning it down costs you nothing at the meter. It costs you accuracy, and only accuracy.

On Muse Spark 1.3 the same ladder exists on the vendor's side, and every notch you turn up is billed. At $4.25 per million output tokens with reasoning tokens counted as output, a workflow that runs at maximum effort is paying for the model's deliberation in the most literal way available. This is the one place where the open model's economics are not just defensible but genuinely better: a batch job that can afford to think harder is free to do so, and a batch job that does not need to think hard is free to stop.

Where the gap disappears

Three situations flatten the nine-point difference to nothing.

Fine-tuning is the first and largest. Muse Spark 1.3 cannot be tuned on your data at any price. Muse Spark Open can, with TorchTitan support that Meta shipped alongside the weights. A model tuned on your codebase, your schema or your domain vocabulary is not the same model as the one that scored 39, in either direction — and for a narrow task, the tuned version of a weaker general model routinely beats the untuned version of a stronger one. If your problem is specific, the index gap is not the relevant measurement.

Data residency is the second. If the corpus cannot leave your network, the comparison is not between 39 and 48; it is between 39 and not having a model. That is not a close call and never was.

Amortised batch volume is the third, with a caveat. A pipeline whose token consumption is large, steady and predictable can justify the hardware. A pipeline that is bursty cannot, because self-hosting converts a variable cost into a fixed one and then charges you for the idle hours regardless. Note also that Muse Spark 1.3's cache discount makes the hosted model unusually cheap precisely for the repeated-context pattern that large batch pipelines produce, so the self-hosting case is weaker than a naive tokens-times-price calculation suggests.

Where the gap becomes a wall

Two workloads should not go near the open model.

Interactive agents are the first. Muse Spark 1.3 measures 127.7 tokens per second, and an agent loop multiplies that number by the number of round-trips per task. Muse Spark Open has no published speed, and the structural expectation is that an 816B model running on desktop memory bandwidth will be far behind. If a human is waiting, the nine index points are irrelevant and the throughput gap is everything.

Bursty frontier reasoning is the second, for a subtler reason. The workloads that need 48 points of reasoning capability are, almost by definition, the ones where a failure is expensive and the reasoning chain is long. Running those on a model that is a tier below, at unknown throughput, to save a per-token bill, is trading a small predictable cost for a large unpredictable one. That is the trade the third team at the top of this piece should refuse.

The team that wants both

Most organisations that end up running both do it because the workloads genuinely differ: interactive paths need speed, batch paths need cost, and only some paths need to stay on-premises. The failure mode is not the architecture — it is that the split lives in a document instead of in configuration, and a developer copies a model string from a tutorial.

A routing layer is the mechanism that fixes that. OrcaRouter routes Muse Spark 1.2 at Meta's list price with 0% markup and passes the vendor's list price through, so a price cut on the hosted side reaches your key the same day with no repricing work. It does not host Muse Spark Open, and it should not be read as doing so: a model whose entire proposition is that you can hold the weights is a model you run on your own hardware. What the router contributes is the interactive half — Muse Spark 1.3 is not currently routed by OrcaRouter either, so the practical shape today is to call the hosted tier wherever you already call it and let one API, with automatic failover and a routing DSL, keep the 200+ models you can reach under a single key while your own hardware handles what has to stay home.

Choose by workload, not by licence

Take Muse Spark Open if your task is narrow enough to tune, if the data cannot leave, if the workload is large and steady, and above all if the model can think at whatever effort level it likes without the meter running. Those are the four conditions, and they are usually correlated — pipelines are narrow, batchy and stationary.

Take Muse Spark 1.3 if a human waits on the output, if the reasoning is genuinely hard, if the work is bursty, or if you need a confirmed multimodal roster — 1.3's text, image and video input is independently listed, whereas Muse Spark Open's wider audio and video claims are vendor-asserted and independently measured as text-and-image only.

What you should not do is take the open weights because they are free. The nine points between 39 and 48 are real, they are concentrated in exactly the agentic and reasoning work that makes these models useful, and no amount of local inference makes them back. Muse Spark Open is the right answer to a specific set of questions. It is not a discount on Muse Spark 1.3.

Muse Spark Open is downloadable under the Muse Model License; Muse Spark 1.3 is available per token at $1.25 input and $4.25 output, with the cache discount applied at the vendor's own rates.