Muse Wonder vs Kimi K4: One Model You Can Buy, One Date You Can't

Any page titled Muse Wonder vs K​imi K4 owes you one sentence before the comparison: K​imi K4 does not exist as a product. M​oonshot AI has published no model card, no repository, no API identifier and no price row for it, and its own platform documentation still tops out at Kimi K3.1 Pro. Muse Wonder, by contrast, has been on sale since 12 October 2026 at $4 per million input tokens and $20 per million output, and as of 9 November it is also served from G​oogle Cloud in preview. So this is not a head-to-head between two shippable models. It is a comparison between a model you can call today and a launch window that shifted, twice, in the past week — and the second shift is the one worth reading about.

That shift happened on 7 and 8 November, when Chinese technology outlets reported a K​imi K4 specification: roughly 5.4 trillion parameters, arriving in the second quarter of next year. That is a leak, not an announcement. Moonshot has said nothing, no artefact corroborates the number, and the sourcing behind it is the same thin trail of chip-procurement reporting that has surrounded K4 since July. But it is the first time a date has attached to the name, and a date changes the shape of the decision even when the date is wrong.

What the leak does and does not tell you

The reported figure is 5.4 trillion parameters, which would put K4 at roughly double the 2.8 trillion of Kimi K3 and K3.1 Pro, and the reported timing is the second quarter of 2027. Both numbers are unverified. Neither appears in any Moonshot document, and the K3.1 Pro model card — which is public, downloadable and inspectable — describes a 2.8-trillion-parameter mixture-of-experts with 104 billion active parameters. There is no published K4 card to compare against.

What the leak does establish is intent. A lab that is training a model roughly twice the size of its current flagship, on a timeline measured in quarters, is telling you that K3.1 Pro is not the end of the line. What it does not establish is anything you can budget against:

• Parameters — 5.4 trillion reported, unconfirmed; K3.1 Pro's 2.8 trillion total and 104 billion active are published and independently checkable

• Date — second quarter of 2027 reported by Chinese outlets on 7–8 November 2026; Moonshot has published no date at all

• Price — nothing reported, and no basis for a guess, because the two previous K​imi generations did not hold their rate cards steady across releases

• Licence — nothing reported; K3.1 Pro ships under a named licence with a commercial-use threshold, and whether K4 follows that pattern is unknown

• Evidence trail — chip-procurement reporting and an internal banner photographed in July, which is where the name came from in the first place; nothing since has been a primary source

If you are weighing a migration, that list is the whole case for waiting, and it is thin. A parameter count is not a capability, and a quarter is not a date you can plan a release around.

What Muse Wonder actually is, now that it is easier to reach

Muse Wonder is the frontier model M​eta released on 12 October 2026, and the G​oogle Cloud preview that went up on 9 November is the most consequential thing that has happened to it since launch. The model did not change. The route to it did.

• Identity — muse-wonder-1.0, a proprietary M​eta checkpoint served on the M​eta Model API, inside Muse Code, and now through G​oogle's partner-model endpoint under publishers/m​eta on the global location

• Context and modalities — 1,048,576 tokens in, with text, image, audio and video accepted and text returned; output ceiling just under a million tokens

• Reasoning control — always on, with an effort ladder from minimal to max and max as the default, which matters for cost because nothing about this model is cheap per token

• Price — $4.00 input, $20.00 output and $0.20 cached input per million, flat at any context length; web search grounding adds $2.50 per 1,000 queries

• The variant you probably cannot have — Muse Wonder Full scores higher and is sold only inside the two most expensive Muse Code subscription tiers, not per token and not on any API

• Independent reading — Artificial Analysis puts standard Muse Wonder at 44 on its Intelligence Index v5.0.1, ranked 32nd of 74 models, with a $6.95 cost per completed index task and 126 output tokens per second

The G​oogle Cloud preview has terms worth knowing before you route production traffic at it: a global endpoint with no regional alternative, a stateless Responses API that does not support store: true, and M​eta's own terms rather than G​oogle's as the contract of record. For a prototype that is fine. For a workload with residency requirements, it is a question to answer first.

The comparison that can actually be made

With K4 unavailable, the K​imi side of any honest comparison is Kimi K3.1 Pro — the 2.8-trillion-parameter model that went on sale on 23 October 2026 and had its weights published on 4 November as a 1.56-terabyte download across 96 shards. Against Muse Wonder it is a genuinely close call on quality and a wide-open one on economics and speed, and the two models fail in opposite directions.

• Intelligence Index — Muse Wonder 44 against Kimi K3.1 Pro 46, per Artificial Analysis v5.0.1; a two-point gap that is inside the noise of most real evaluations and should not decide anything on its own

• Cost per completed index task — $6.95 for Muse Wonder against $3.68 for Kimi K3.1 Pro, which is the number that survives contact with an invoice; the cheaper model is also the higher-scoring one here

• Output speed — 126 tokens per second for Muse Wonder against 32 for Kimi K3.1 Pro, a factor of four that no benchmark row captures

• Time to first answer token — about 20 seconds for Muse Wonder against roughly 155 seconds for Kimi K3.1 Pro, because the K​imi model thinks before it speaks and the bill is charged in wall-clock patience

• Token price — $4.00 and $20.00 per million for Muse Wonder against $2.00 and $10.00 for Kimi K3.1 Pro, with a 90 per cent cache discount on the K​imi side

• Weights — Muse Wonder is closed and not downloadable in any form; Kimi K3.1 Pro is a public checkpoint you can host yourself, subject to a licence with a commercial-use threshold

• Verbosity — both burn around 113 million output tokens on the index run, which is unusually matched and means the per-task gap is a rate difference rather than a chattiness difference

Put plainly: Kimi K3.1 Pro halves the bill and triples the wait, and scores two points higher while doing it. Muse Wonder answers four times faster and costs about twice as much per task. Neither of those is a defect. They are different bets about what the model is for — an overnight batch job and an interactive agent have almost nothing in common as purchasing decisions.

What would have to be true for waiting on K4 to make sense

The case for waiting is not irrational, but it is narrower than the search volume around K4 suggests. Four things would have to hold at once, and three of them are currently unknown:

• The date has to be real — a reported second quarter of 2027 is not a date until Moonshot publishes one, and the current flagship shipped roughly two months after the previous generation's API launch, so the cadence is not slow enough to make waiting obviously cheap

• The weights have to follow — the pattern Moonshot has set twice now is API first, weights later, and a K4 released closed would break the single strongest reason to prefer the K​imi line for anyone who needs to self-host

• The licence has to stay usable — K3.1 Pro's licence already carries a commercial-use threshold rather than an outright grant, and a stricter K4 licence would erase the self-hosting advantage entirely

• The price has to land under the current model — with no reported rate card, that is an assumption, and K3.1 Pro at $2 and $10 is already the cheaper side of this comparison

Set against that, the cost of waiting is concrete: you spend a quarter on the model you have, or you defer a decision that a two-point benchmark gap cannot justify deferring. The one scenario where waiting is clearly right is a team that needs a downloadable checkpoint and expects K4 to be one. That team should build on K3.1 Pro now, because the migration path from K3.1 Pro to K4 — if K4 ever ships with weights — will be shorter than the path from nothing.

Running the half of this comparison that exists

OrcaRouter does not carry Muse Wonder, so if Wonder is what your workload needs, M​eta's own API and the G​oogle Cloud preview are your two routes and there is no third. What we do carry is Kimi K3.1 Pro — k​imi/kimi-k3.1-pro, at Moonshot's own $2.00 and $10.00 per million tokens with zero markup, in a catalogue of more than 200 models behind a single key.

That matters for a specific reason in this matchup. The realistic 2027 plan for anyone reading a K4 leak is not "switch to K4 on day one"; it is "keep a working K​imi integration until K4 lands, then test it against what is already running." Doing that on one API means the swap is a model-ID change rather than a new vendor contract, and automatic failover means a preview endpoint or a launch-day capacity crunch does not take the workload down while you evaluate the successor.

The same routing layer is what makes the Muse side testable without a bet: point a fraction of traffic at Muse Wonder through M​eta's API, compare completed-task cost against the K​imi baseline on your own work rather than on an index, and let the numbers decide. The two-point Intelligence Index gap will not survive contact with your particular prompts; the four-fold speed difference almost certainly will.

Where this lands

Muse Wonder wins this comparison by default, in the only sense that matters commercially: it exists, it is priced, it answers in twenty seconds, and since 9 November it is available through a cloud contract many enterprises already hold. K​imi K4 wins nothing yet, because there is nothing to win with — a reported parameter count and a quarter are a forecast, not a product.

The more useful comparison, and the one worth acting on, is Muse Wonder against Kimi K3.1 Pro, and it comes out close to a coin flip that your workload has to break. If latency is the constraint, Wonder. If cost per completed task is the constraint, K3.1 Pro — by roughly half. If you need to own the checkpoint, only one of the two is a candidate at all, and the K4 question is really a question about whether that stays true in 2027.

Watch three things and nothing else: a Moonshot model card with a K4 identifier on it, a weight release following it, and whether the licence is still one a company can live under. Until the first of those appears, "Muse Wonder vs K​imi K4" is a search with a real answer, and the answer is that you are comparing a model to a rumour.