Muse Wonder vs Kimi K3: The Downloadable Model That Is Still Not Free to Resell

Kimi K3 is the open-weight answer to everything Muse Wonder represents, and the interesting part is that its licence does not quite let you do what "open weight" usually implies. Moonshot AI released Kimi K3 on July 16, 2026: 2.8 trillion parameters with 104 billion active, FP8 weights ungated on Hugging Face, a 1,048,576-token context, and a rate card of $3.00 per million input tokens and $15.00 per million output. Muse Wonder arrived on October 12, 2026 from Meta Superintelligence Labs at $4.00 and $20.00 across the same sized window, with no weights at all. On the independent index the gap is eight points — 52 for Muse Wonder (max) against 44 for Kimi K3 (max) — and about $4.52 per index task against $2.00. But the clause that decides which one your business can actually build on is buried in Moonshot's licence, not in either score.

What the licence does, and does not, grant

Kimi K3 ships under the Kimi K3 License, which Moonshot describes as derived from MIT. It is open-weight, a term Moonshot uses deliberately and does not upgrade to open-source, and the distinction is load-bearing because three conditions ride along with the download.

• Copies must carry the notice and comply with law.

• A model-as-a-service operator earning more than $20 million over any consecutive twelve months needs a separate commercial agreement with Moonshot.

• A product with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display "Kimi K3" prominently.

• Internal use, Moonshot's own products and certified inference partners are exempt from the second and third conditions.

Condition two is why Kimi K3's arrival on Amazon Bedrock in September had to be negotiated rather than simply uploaded, and it is the clause that matters to anyone planning to wrap the model in a product. Downloading the checkpoint, fine-tuning it, running it inside your own company — all permitted. Standing up an inference business on it above the revenue line — not permitted without a contract. Muse Wonder's proprietary weights have no licence to read because there are no weights; the trade you make there is different, not worse.

The practical footprint is worth stating plainly too, because "open weight" invites a picture of a workstation. 2.8 trillion parameters at FP8 is a multi-node deployment — eight H100- or H200-class accelerators at minimum, and in practice more — which puts self-hosting out of reach for anyone without a cluster. The download exists. The ability to use it cheaply does not follow automatically.

Cost per task, where the two separate

Both models are cheap per token and neither is cheap per completed job, and the ordering is the same in both views for once:

• List rate — Kimi K3 $3.00 in / $15.00 out / $0.30 cached per million; Muse Wonder $4.00 in / $20.00 out / $0.20 cached

• Cost per intelligence index task — Kimi K3 about $2.00; Muse Wonder about $4.52

• Output tokens on the index run — Kimi K3 about 160 million; Muse Wonder about 130 million

• Intelligence Index — Muse Wonder (max) 52, ninth of 219; Kimi K3 (max) 44

• Context window — both 1,048,576 tokens

Kimi K3 wins the cost argument by 2.3×, and it wins it despite being the more verbose of the two, because its output rate is $15 against Meta's $20. The eight-point index gap is the price of that saving. Both figures come from Artificial Analysis on the same revision of its index, so the two per-task numbers are directly comparable in a way two vendor scoreboards never are.

Muse Wonder's one unambiguous pricing advantage is the cache read: $0.20 per million against Kimi K3's $0.30, on a context window of identical size. For an agent loop that re-sends a large fixed prefix on every step, that difference accumulates on the largest token category in the request and is not offset by anything on the output side.

Where Kimi K3's composite score stops predicting

An eight-point index gap is not spread evenly, and on the sub-scores the two models stop looking like neighbours.

• Terminal-Bench 4.0 — Kimi K3 0.126; Muse Wonder not published on the independent harness, but Meta's own unreproduced table claims 85.6 against 81.4 for GPT-6 Astra

• AA-Omniscience — Kimi K3 19.7; GPT-6.1 Sol 41.5, for scale on the same index

• Long-context retrieval — Kimi K3 approximately 0.887 on AA-LCR v1.1, one of the strongest readings on the board

• Output speed — Kimi K3 about 44 tokens per second with a median time to first token near 8.0 seconds; Muse Wonder not yet published

Read those together and a shape appears. Kimi K3 is excellent at reading an enormous input carefully and answering about it, and weak at executing a plan across many small steps with checks in between. Its long-context retrieval result is better than most frontier models', while its terminal-execution result is near the floor. That is consistent with the community reading of the model — strong when a task can be answered in one long act of reasoning, fragile when it must be executed stepwise — and with the architectural story, since Kimi Delta Attention is a hybrid linear-attention design that trades some long-context fidelity and some interactive latency for efficiency at scale.

Muse Wonder's claims point the other way. Meta's launch table concentrates its wins in Terminal-Bench, SWE-Marathon and computer-use suites, and its own AutomationBench row is its weakest relative to the Claude and GPT-6.1 Sol columns in the same table. If that shape is real, the two models are complements: K3 for reading, Muse Wonder for doing. The caveat is unavoidable — Meta's Terminal-Bench number is vendor-reported and unreproduced, and no harness has yet put both models through the same agent suite.

The cost of leaving

The strongest argument for Kimi K3 was never the price. It is that the checkpoint keeps working if Moonshot changes its terms, and that a team with the hardware can pin a version and stop caring what happens to the vendor's rate card, capacity or roadmap.

Two caveats sit under that argument, and both are documented rather than speculative. Moonshot has had capacity trouble: the first-party API has been described as having constant capacity issues, and the company paused new consumer subscriptions shortly after launch. Second, the licence's commercial-agreement condition means the escape hatch is a real escape hatch only for internal use — the moment the model becomes part of a service you sell at scale, you are negotiating with Moonshot anyway. Muse Wonder has no such concerns and no such option: Meta's rate card is the rate card, and the only hedge is to have a second model ready.

Which of them you can call today

Kimi K3 is in the OrcaRouter catalogue as kimi/kimi-k3, dated 2026-07-15, at Moonshot's own $3.00 input, $0.30 cache-read and $15.00 output per million tokens across the full 1,048,576-token window, behind the same OpenAI-compatible endpoint as everything else. Provider list prices pass through unchanged, so a Moonshot rate change lands here the same day, and an automatic-failover rule means a capacity squeeze on one upstream provider moves the request instead of dropping it — which is worth something specific with this model, given that capacity is the documented weak point.

Muse Wonder is not in our catalogue, and the page says so rather than implying otherwise: Meta's Model API and Muse Code are where it runs. What we do carry from the same lab is meta/muse-spark-1.2, the larger Meta checkpoint, at $1.25 input and $4.25 output per million across a 1,048,576-token window with $0.15 cache reads.

The reason that arrangement is useful for this particular decision is that the two halves of this comparison are decided by different things — K3 by whether your workload executes or merely reads, Muse Wonder by whether you need speech and video as first-class inputs and a 943,718-token output ceiling. Those are testable questions, and one of them is a model-name change away from being tested.

Choosing

• Kimi K3 when the job is comprehension: long documents, large codebases, retrieval over a million tokens, and anything where a single long act of reasoning produces the answer. It is cheaper per task by more than a factor of two, its long-context retrieval is among the best measured, and the weights give you an exit that Muse Wonder cannot offer at any price.

• Kimi K3 when latency does not matter. Forty-four tokens per second and an eight-second median time to first token is a poor fit for an interactive surface and a fine fit for an overnight batch.

• Muse Wonder when the task has to be executed rather than answered — terminal work, computer use, long-horizon agents with tool calls — provided you treat Meta's benchmark table as a claim rather than a measurement.

• Muse Wonder when the input is speech or video, or the output has to exceed 450,000 tokens in one response. Neither is available on the other side.

• Muse Wonder when you have already decided the eight index points are worth 2.3× the run rate. That is a legitimate call for a frontier-parity workload and a bad one for high-volume text processing.

The licence clause is the part most comparisons miss, and it is the part that will matter in eighteen months. Kimi K3 gives you a copy of the model and a contract condition; Muse Wonder gives you neither, and charges more per task for the privilege. Whether that is a bad deal depends entirely on whether you were ever going to run the thing yourself.