Muse Spark vs Muse Spark 1.3: One of These Names Is a Version, and That Is the Whole Answer

Typing Muse Spark against Muse Spark 1.3 into a search box implies two products, and the first useful thing anyone can tell you is that there is only one. Meta Superintelligence Labs introduced Muse Spark on April 8, 2026, and Muse Spark 1.3 — released September 2, 2026, with its max reasoning level opened to developers on September 4 — is what that name now points at. Meta's own model page settles it in three sentences: Muse Spark 1.3 is "the latest version" and is recommended for new work, Muse Spark 1.2 is "the previous version," and Muse Spark 1.1 is "the original version." So this is not a matchup. It is the version question that the Muse Spark name has been hiding since the family started re-pointing, and the answer changes depending on which of the earlier checkpoints you were actually comparing against.

What each name denotes

The unversioned name is a family and a moving target. Meta shipped the original Muse Spark in April 2026 as a natively multimodal reasoning model with a private API preview, then began issuing numbered checkpoints: Muse Spark 1.1, which our own catalogue lists with a release date of July 16, 2026; Muse Spark 1.2 on August 5, 2026; and Muse Spark 1.3 on September 2, 2026. Independent listings agree with those latter two dates. There is one wrinkle worth flagging rather than smoothing over: Meta's documentation calls 1.1 "the original version" while the family introduction was in April, and the gap between those two statements is not explained on any page we could read. Treat the April date as the family's introduction and July as the first checkpoint that carries a number, and do not treat either as a launch event this page is reporting.

All three numbered versions share the same modality row and the same 1,048,576-token context window, which Meta states directly: they "share the same modalities and context window, and differ only by capability." That sentence is the most practically useful thing on the page. It means a version bump in this family does not change your request shape. Same inputs — text, image, video, audio and PDF — same text-only output, same ceiling. What changes is the quality of the answer, and the price does not change at all: $1.25 per million input tokens, $0.15 per million cached input and $4.25 per million output tokens, identical across 1.1, 1.2 and 1.3 on Meta's Standard tier.

The upgrade arithmetic, from Meta's own table

Meta publishes a single benchmark table that puts Muse Spark 1.3 running max reasoning beside Muse Spark 1.2 running xhigh across a dozen agentic and long-context evaluations, and reading it row by row produces a more interesting picture than the version number suggests. Every figure below is vendor-reported, produced inside Meta's own harness, and unreproduced by us.

• GDPval-AA v2 — Muse Spark 1.3 (max) 1,754 vs Muse Spark 1.2 (xhigh) 1,615

• Terminal-Bench 2.1 — 88.8 vs 82.9

• OSWorld 2.0 — 66.9 vs 47.6

• DeepSWE v1.1 — 75.4 vs 55.0

• SWEAtlas — 59.4 vs 46.2

• JobBench — 64.9 vs 61.6

• MRCR 256K–512K — 98.5 vs 66.3

• MRCR 512K–1M — 98.1 vs 55.5

The last two rows are the story, and they are not the rows anyone quoted at launch. Long-context retrieval at the 256K–512K band moves 32 points; at the 512K–1M band it moves more than 42. General capability work moves far less: GDPval-AA up 139 points on a scale that had 1.2 at 1,615, and JobBench up 3. The pattern is consistent with what a lab tunes when it is chasing long-horizon agentic work — the model did not get uniformly smarter, it got dramatically better at holding and retrieving information across a context it already had the window for. If your workload lives in the first 100K tokens, the upgrade is real but modest. If it lives beyond that, it is the difference between a model that can use a large context and one that can find things in it.

Two caveats have to sit on those numbers. The comparison is max against xhigh, not max against max, because Meta ran 1.2 at xhigh and 1.3 at max — the settings differ, so part of the gain is reasoning budget rather than checkpoint quality. And 1.2's row in that table is now a deprecated configuration: outside Meta's own table, independent listings flag Muse Spark 1.2 as deprecated and point to 1.3, which is the strongest available signal that the older checkpoint is on its way out.

The independent picture, and the trade nobody mentions

Artificial Analysis is the third party with public listings for both checkpoints, and its numbers disagree with Meta's in an instructive way — not about direction, but about cost.

• Muse Spark 1.3 (Max) — Intelligence Index 48, rank 24 of 217; output speed 121.9 tokens/second; $1.60 per completed Index task; 170M output tokens generated during the evaluation

• Muse Spark 1.2 (Xhigh) — Intelligence Index 40, rank 59 of 217; output speed 361.4 tokens/second, ranked first of 186 models; $0.97 per completed Index task; 130M output tokens; time to first token 12.67 seconds

Read the two lists against each other and the upgrade looks different from how it reads in Meta's table. Muse Spark 1.3 is eight index points stronger and roughly three times slower — 121.9 tokens per second against 361.4, where the older checkpoint was the fastest model of any kind on that board. Cost per completed task rose from $0.97 to $1.60 even though the per-token price is identical, because the newer model generates 40 million more tokens to finish the same evaluation. Nothing about the vendor's rate card changed. What changed is how much the model says.

That is the trade the version number conceals, and it cuts in a direction most coverage missed. If your application is latency-sensitive or throughput-bound, the previous Muse Spark version may genuinely be the faster product for you, and at the same list price. If it is accuracy-bound on hard problems, 1.3 with max reasoning is the point of the upgrade. These are different purchases dressed as a version bump, and only one of them is what the deprecation notice assumes you want.

What 1.3 added that is not a score

The max reasoning level is the most consequential change in the 1.3 release and it arrived two days late. At the September 2 launch Meta shipped every effort level up to xhigh and held max back pending what it called additional safety testing, sharing it with a limited partner set. On September 4 that testing finished and max became selectable in Muse Code and on the muse-spark-1.3 model id through the Meta Model API. The consequence is that the headline scorecard — the 1,754 on GDPval-AA, the 98.5 on MRCR — was produced at a setting developers could not call for the first two days, and any integration written on launch day may still enumerate reasoning effort only as far as xhigh. Check what your route accepts before assuming max will be honoured.

Meta also sells the 1.3 checkpoint twice, and the difference is not the model. muse-spark-1.3 is labelled "Not used to improve our products" at $1.25 / $0.15 cached / $4.25. muse-spark-1.3-contributor is labelled "Used to improve our products" at $0.10 / $0.002 / $0.20 — roughly an order of magnitude cheaper on input and more than twenty times cheaper on output, in exchange for permission to train on your prompts and completions. Both carry the same 1M-token context window. Contributor describes what you contribute, not what you receive; it is a pricing tier with a data term attached, not an open-weights release.

And there is one regression, stated by Meta itself rather than discovered by a user. Audio understanding in Muse Spark 1.3 is not fully supported, and Meta warns that response quality on requests carrying audio content may be degraded, pointing audio workloads at Muse Spark 1.2. So the current version is not the right answer for every input in the family's own modality row — a case where the newest checkpoint is the wrong one for a payload the older one handles.

Which version you can actually call

Worth separating the version question from the availability question, because they have different answers. All three numbered Muse Spark versions are hosted only: Meta publishes no weights, no licence and no download for any of them, and the Contributor variants are discounted hosted tiers rather than releases. The open weights in this family belong to Muse Spark Open, the 816B sparse mixture-of-experts model Meta published on October 15, 2026, and to Muse Glimmer, the 30B Apache 2.0 model from August 10, 2026. Neither is a Muse Spark version, and neither is a substitute for one.

Our own catalogue reflects the older end of the line, and we would rather say that plainly than let a version number imply coverage we do not have. OrcaRouter routes Muse Spark 1.1 and Muse Spark 1.2, both at Meta's list price with 0% markup passed through, both on the family's 1,048,576-token window. Muse Spark 1.3 is not on our catalogue today, and neither Contributor variant is. That makes the practical upgrade path a two-step: 1.2 is what you can route to now, 1.3 is what Meta recommends for new work, and the day it lands on a provider we route the change arrives on the same key rather than in your integration code.

The version question is also where a router earns its keep in a way that is specific to this family. Three checkpoints, one vendor, one endpoint, no weights — the entire line runs through a single API operated by a single company, which is the concentration risk of building against it. Routing 1.1 and 1.2 behind one OpenAI-compatible endpoint means a failed long-context call can fail over to the other checkpoint instead of failing the request, and the routing DSL lets you split production traffic between them by configuration, so you can measure whether 1.2's speed is worth more to you than 1.3's accuracy on your own prompts rather than on someone's index. That is the only version comparison that settles the question for your workload.

The answer, depending on which question you arrived with

If you wanted to know whether Muse Spark and Muse Spark 1.3 are different models: they are not. 1.3 is what the name means now, and Meta labels it the latest and recommends it for new work. If you wanted to know what you get by moving from the previous version: roughly 32 to 42 points of long-context retrieval, a large jump in agentic computer use and long-horizon coding, and a three-fold drop in output speed at the same per-token price. If you wanted to know what it costs: the rate card has not moved across three versions, which tells you Meta is competing on capability per dollar rather than on the dollar.

What none of these numbers can tell you is which side of that trade your own workload sits on, and the honest position is that the vendor's deprecation notice assumes an answer for you. The version that is three times faster at the same price is still the only one of the two you can reach through a router today, which makes this an unusually cheap experiment to run before 1.3 arrives.

What the numbers cannot tell you is whether any of it has gone wrong in the field: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.