Muse Spark vs Kimi K3.1: The Dot in the Model Name Is the Whole Story

Two model names, two different kinds of missing information. Muse Spark 1.3 is a shipped, measured, callable model: Meta Superintelligence Labs' current checkpoint, generally available since September 2, 2026 at $1.25 per million input tokens and $4.25 per million output, scoring 48 on Artificial Analysis' Intelligence Index at its max reasoning setting. Kimi K3.1 is not a model at all. There is no card, no weight repository, no rate-card row and no API identifier carrying that number anywhere on the vendor's own surfaces, and there has not been since a community reading of a deleted teaser pointed at it in September. What the dot resolves to is the line it abbreviates — Kimi K3, the 2.8-trillion-parameter open-weight flagship the lab shipped on July 16, 2026 — which is callable today at $3.00 per million input tokens and $15.00 per million output. So the useful version of this comparison is Muse Spark 1.3 against Kimi K3, with the gap between the two of them narrow enough that the naming mistake stops being pedantic and starts being expensive.

Where the dot came from, and what it is worth

The strongest signal that a Kimi K3.1 exists is also the weakest kind of evidence: a post attributed to Kimi's official account on or around Friday, September 18, 2026, publishing a long string of digits and deleting it shortly afterward. The digits are the decimal expansion of pi with its opening 3.1 removed — pi is 3.1415926535, and the posted string began 41592653 — and community threads read that as a version number almost unanimously. It is a vendor-side artifact and it is unverifiable, because the post is gone. Reading "K3.1" out of it is an inference, not a statement Moonshot has made.

Four weeks on, nothing has been added to it. The August window that leakers had pointed at came and went with no launch. The anonymous arena entry that some read as Kimi K3.1 earlier in the summer was, on the available evidence, probably not Kimi. The newest signal in the story is a third-party guess that Kimi K3.1 and two Z.ai model numbers land in the same week, which is one person's read on another person's vibes rather than a leak, and the week it names is nearly over without any of the three appearing. Moonshot's documentation, its pricing page, its Hugging Face organisation and its blog all use one name for one model — Kimi K3 — and none of them uses a dotted suffix. The dot is a search-box habit, not a version.

That matters here for a reason that has nothing to do with tidiness. Readers type the dot because they have already read something about this line and are coming back for it. Arriving at a comparison built on a model that does not exist, they would be comparing Muse Spark 1.3 against an inference.

The matchup that exists: 1.3 against K3

Both are flagship-class systems with a million-token window and a stated multimodal input row, and both were sold on long-horizon agentic work — Meta's line for 1.3 is "trained for agentic workflows and optimized for competitive coding performance," Moonshot positions K3 for "long-horizon coding and end-to-end knowledge work." The measured difference between them is four index points, which is exactly the margin where the other columns decide the answer.

• Intelligence Index — Muse Spark 1.3 at max effort 48, rank 24 of 217, against Kimi K3 at max effort 44, ranked third on the panel. Four points, Artificial Analysis' own hosted measurement on both sides.

• Output speed — Muse Spark 1.3 measured at 121.9 tokens per second, one of the faster flagships on that board; Kimi K3 is measured at 44 tokens per second and flagged as notably slow, roughly a 2.8x gap in wall-clock throughput.

• Cost per completed index task — $1.60 on Muse Spark 1.3 against $2.00 on Kimi K3. Note that this is the reverse of the per-token comparison, and the reason is verbosity: Kimi K3 generated about 160 million output tokens across the index run against a median of 130 million, while Muse Spark 1.3 generated roughly 170 million against a median of 87 million. Both are measured as talkative; Kimi K3's price per token is simply higher to begin with.

• Rate card — Muse Spark 1.3 at $1.25 and $4.25 per million tokens, Kimi K3 at $3.00 and $15.00. Kimi K3 carries a 90% cache-read discount, so on cache-heavy multi-turn workloads the effective spread narrows sharply; Moonshot bills cache writes separately, and at the one-hour retention tier they are more expensive than a fresh input token.

• Context and output — both 1,048,576 tokens of context. Moonshot publishes the same figure on its rate card, so the two are matched on the dimension both vendors lead with.

• Sampling controls — Kimi K3 fixes temperature at 1.0, top_p at 0.95, n at 1 and both penalties at 0, and returns an error if you pass anything else. Depth is the only dial, through a top-level reasoning_effort field taking low, high or max and defaulting to max, and changing it mid-conversation invalidates prefix-cache hits. Muse Spark 1.3 exposes a reasoning-effort setting too, and Meta's tooling notes that turning reasoning off is not possible. Practically, K3 removes an entire class of tuning from the job and hands you one lever; 1.3 leaves the ordinary sampling parameters in play.

The shape of the trade is legible without interpretation. Muse Spark 1.3 is faster, cheaper per finished task and four points higher on the index. Kimi K3's case is not built on any of those — it is built on the weights.

Weights, and which side of the licence line each model sits on

This is the column that decides the matchup for a lot of teams, and it runs the opposite way from the scoreboard. Kimi K3 ships open weights in the moonshotai/Kimi-K3 repository under the Kimi K3 License, and Moonshot describes it as the first open model in the three-trillion-parameter class: 2.8 trillion total parameters with 104 billion activated per token across 896 routed experts, 16 firing per token plus 2 shared experts, 93 layers of which one is dense. It is published with quantisation-aware training from supervised fine-tuning onward, MXFP4 weights and MXFP8 activations, which is Moonshot's argument for broad hardware compatibility. Everything in that paragraph is vendor-stated and not independently reproduced, and the distinction is worth keeping, because Moonshot is the party that both trained the model and reports what it scores.

Muse Spark 1.3 has none of that. It is a served checkpoint, and the Muse Spark Open weights Meta published on October 15, 2026 are a different model with a different score — 39 on the same index at its xhigh setting, ninth of 118 open-weight entries above 150 billion parameters. So the honest statement of the weights position is not "Meta opened Muse Spark," it is that Meta now publishes an open artifact next to a closed line, and if open weights are the reason you came, the K3 side of this page delivers them for the model in the comparison while the Meta side does not.

Where OrcaRouter sits in it

Kimi K3 is in OrcaRouter's catalogue at Moonshot's published rates — $3.00 per million input tokens and $15.00 per million output, with the model's 1M-token context written on the page — and it routes on the same key as Muse Spark 1.2 and Muse Spark 1.1, the two Muse Spark checkpoints we serve at Meta's list price of $1.25 and $4.25 with 0% markup, because the provider's list price is passed through rather than marked up. That combination is the cheap way to settle a four-point index question that no scoreboard can answer for your traffic: the routing DSL splits requests by rule, so you can send the latency-sensitive path to the faster model and the batch path to the one whose weights you want to standardise on, and automatic failover behind one endpoint keeps either from being a single provider bet. Muse Spark 1.3 itself is not on the catalogue — the served checkpoints are 1.2 and 1.1 — and the benchmarks quoted above are 1.3's, so the routing comparison and the measured comparison are not the same pair of models. That gap is worth knowing before you plan around it.

What to do with a name that has no product behind it

If you arrived from a link about Kimi K3.1, the correct next step is not to compare anything — it is to decide which shipped model you are actually choosing between. Against Kimi K3, Muse Spark 1.3 wins on index, speed and cost per completed task, and loses on weights and on the licence freedom that comes with them. Against the wider Kimi line, the smaller K2.6 and K2.7 Code models sit in a different price class entirely and are the right comparison for anything that does not need three trillion parameters. Nothing in this page is a reason to wait for a dotted suffix: on the current record there is no card, no date and no endpoint behind it, and the only firm thing anyone knows about it is that the post that pointed at it was deleted.

What would change the picture is narrow. Moonshot publishes a K3.1 model card or a rate-card row, and this becomes a real two-model page with a second column of measured numbers. Artificial Analysis adds a K3.1 listing, and the four-point gap is testable rather than inferred. Or Meta re-points the Muse Spark line at a 1.4 checkpoint, which moves the Meta side for reasons unrelated to Moonshot. Until at least one of those lands, the factual version of this matchup is Muse Spark 1.3 against Kimi K3, separated by four points of measured intelligence, roughly 2.8x of throughput and a licence boundary — and the dot is a search term, not a product.

None of this is a reason to avoid the model, only a reason to know the history: the documented AI incidents and agent failures behind 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.