Muse Wonder vs Qwen 4 Max: One Shipped Three Weeks Ago and One Is Still in Training

Meta Muse Wonder has an API identifier, a token price, a context window, two checkpoints and a live endpoint that has been serving requests since October 12. Qwen 4 Max has a name, a stage appearance and a sentence in a press release. The vendor's own written statement from its Yunqi/Apsara conference in Hangzhou on 22 September 2026 says only that Qwen 4 is in training — it carries no price, no window, no parameter count, no licence, no benchmark and no identifier for any tier, and the family's LLM lead's stage preview of Qwen 4 Max, Qwen 4 Flash, Qwen 4 Plus and Qwen 4 27B was not converted into a specification anywhere on the vendor's surfaces in the three weeks since. So this is not the usual head-to-head, and the useful version of the question is a timing one: what can you run today, and what evidence would tell you Qwen 4 Max is close enough to wait for.

What the vendor actually put in writing

The vendor's own material is one press release, and it is worth quoting precisely because everything else in circulation is inference from it.

• Qwen 4 — described as the next-generation model and as currently in training

• Qwen 4.5 and Qwen 5 — named as later series, with a projection of scaling to 5 to 10 trillion parameters

• Qwen 4 Max, Flash, Plus and 27B — shown on stage as a four-tier roster, not documented in the release

• Architecture — described as a new generation and described as being trained, which is not the same as being finished

• Primary framing — agentic and tool-using work rather than chat

• Availability — none. No tier has a release date, a price, a context window, a modality list or a published score

Two corrections belong here because both are circulating as fact. The 5-to-10-trillion-parameter figure attaches to the Qwen 4.5 and Qwen 5 series in the vendor's own sentence, not to Qwen 4 and not to Qwen 4 Max; any page printing it as a Qwen 4 Max specification has moved it a generation. And the framework support that has landed in vLLM, SGLang and Transformers over the past two months for something called "Qwen4-Exp" is not Qwen 4 Max. That name belongs to the experimental architecture whose public release is Qwen3.8-Flash-Next, which shipped on 26 August; its own model card is the one that calls itself a preview of the design that will underpin Qwen 4.

What that card does tell you about the family is more concrete than the tier roster, and it is worth holding as a shape rather than a spec sheet: roughly 125B main parameters with about 6B activated, a 51B n-gram embedding, 262,144 native context extensible toward a million, QSA sparse attention, gated residuals, and a training-cost figure described as roughly 90% below its predecessor. That is the experimental sibling's design, attributed by the vendor to that model and explicitly framed as what comes next — not a Qwen 4 Max datasheet, and it should not be read as one.

What the shipping side of the comparison looks like

Muse Wonder is the contrast in every dimension where Qwen 4 Max has nothing yet, so it is worth stating what is genuinely on the table from Meta.

• Availability — generally available since October 12, 2026 as muse-wonder-1.0 on the Meta Model API and inside Muse Code

• Price — $4.00 per million input tokens, $0.20 cached, $20.00 per million output, flat across the window

• Context — 1,048,576 tokens, with a documented maximum output of 943,718

• Reasoning — always on, effort from max down to minimal, default medium, reasoning tokens billed as output

• Second checkpoint — muse-wonder-full-1.0, scoring higher, included with a Muse Code subscription rather than sold per token, on plans published at $50 and $200 a month

• Independent measurement — Artificial Analysis reads Muse Wonder (max) at 52 on its Intelligence Index, ranked ninth of 219 models, about $4.52 per index task and about 130M output tokens on that run, against a board median near 85M

• Vendor claims — Meta's own table reports the Full checkpoint at 75.3 against Muse Wonder's 30.8 on JobBench, a 44-point spread between two checkpoints of the same launch, produced in Meta's harness and unreproduced by anyone

Set that beside a Qwen 4 Max column and the honest rendering of the second column is a row of blanks. That is not a criticism of the model; it is what a tier looks like three weeks after a conference appearance and before a launch.

The Qwen you can actually call, and how it compares

If the reason you are here is that you want a Qwen flagship and the 4-series is not available, the one with a rate card is Qwen3.8 Max, generally available since 3 August 2026. It is not the same model and it is not a substitute on capability — but it is the answer to "what do I run this quarter."

• Price — $2.00 per million input tokens, $0.25 cached, $6.00 per million output, against Muse Wonder's $4.00 / $0.20 / $20.00

• Context — 1,000,000 tokens as the vendor publishes it; Artificial Analysis records the evaluated checkpoint at 984k

• Input surface — text, image and video, against Muse Wonder's text, image, speech and video

• Intelligence — 45.4 on the index our catalogue tracks from Artificial Analysis, against Muse Wonder's 52

• Cost per index task — about $5.41, against Muse Wonder's about $4.52, despite the cheaper rate card

• Output tokens on that run — about 190M, against Muse Wonder's about 130M and a board median near 85M

• Speed — about 36 tokens per second on the evaluator's run, described as notably slow

• Weights — Qwen3.8 Max's were published on 12 August 2026 under a custom licence, as Qwen3.8-2.4T-A95B in BF16 and FP8

That last pair of lines is the most interesting thing in this article. Qwen3.8 Max is half Muse Wonder's input price and under a third of its output price, and it still costs more per completed task — about 20% more — because it burns roughly 190 million output tokens where Muse Wonder burns 130 million. It is the same inversion this family has shown before: the cheaper rate card attached to the more expensive job. And the weighted score is seven points lower at the same time. More tokens, lower score, higher bill is the expensive kind of verbosity, and it is not visible on either vendor's pricing page.

Against that, the open weights are a genuine advantage for a class of buyer. Published weights mean a checkpoint you can audit, quantise, fine-tune or freeze at a version you control; they do not mean you can run it locally, since a 2.4-trillion-parameter mixture-of-experts is a datacentre artefact rather than a workstation model. But if your compliance posture requires an exit from hosted inference, neither Muse Wonder nor Qwen 4 Max will give you one on day one, and Qwen3.8 Max already does.

Should you wait for Qwen 4 Max?

The wait-or-buy question has a real answer if you separate the two things people are actually waiting for.

• If you need a frontier model this month, there is nothing to wait for. Qwen 4 Max has no date, and the vendor's language — in training, arriving soon — does not distinguish a Q4 launch from a Q1 one. The models with endpoints today are Muse Wonder, Qwen3.8 Max and the current Claude, GPT-6 and other frontier flagships.

• If what you are waiting for is the 27B open tier, that is a different calculation. Every Qwen generation has shipped its small tier under terms that let people build, and the community response to the 3.8 series — MLX conversions, NVFP4 quantisations, fine-tunes — appeared within days of the weights landing. An announced 27B is a promise of an ecosystem, and it is the tier most likely to be locally runnable.

• If what you are waiting for is a price cut, waiting is not required. The Qwen 4 generation is expected to bring the architecture preview's cost profile with it, but that is a projection from a sibling model's card, not a published Qwen 4 figure. Meanwhile prices on the current generation have been moving downward, and they move on the routing side the day they move on the vendor's.

Three signals would justify updating this position, and all three are checkable surfaces rather than rumours: an API model identifier appearing on the vendor's own platform, a price row on the same page, or a dated model card with a context window. A conference date is not one of them, and neither is a leaderboard appearance, since aggregator listings have been appearing before vendor documentation for this family rather than after it.

Where OrcaRouter sits in this

The practical reason to run this comparison through one endpoint is that the answer here is a timing decision, and timing decisions need measurement rather than a spec sheet.

Qwen3.8 Max is on OrcaRouter at the vendor's own rate — $2.00 per million input at $0.25 for cache reads and $6.00 per million output — with the million-token window intact, over the GPT-compatible format, the Messages format and the native DashScope shape. We take no markup on any of it, so a price change from the vendor is live on our side the same day rather than at the next renewal. That is the Qwen flagship you can wire into an application today, and its cache-read rate being a tenth of its input rate makes it unusually cheap for the long-context, repeated-prefix work the Qwen 4 family is being aimed at.

Muse Wonder is not on our catalogue and we will not imply it is: Meta's Model API and Muse Code are the routes, and the Full checkpoint is not sold per token at all. Muse Spark 1.2 is the Meta model we do carry, at $1.25 in and $4.25 out with $0.15 cache reads across 1,048,576 tokens — the cheap Meta leg if what you need is the family rather than the flagship.

When Qwen 4 Max does get an endpoint and a price, it will appear here at the vendor's published rate with the same pass-through rule, and the model-fusion and routing-DSL paths will let you put it in a panel against Muse Wonder and Qwen3.8 Max without a third contract. Until then the sensible build is the boring one: ship on what exists, keep the cheap Qwen leg in the cascade, and re-run this comparison when a rate card appears.

The position, stated plainly

Meta Muse Wonder is a model you can call, price, benchmark and budget against. Qwen 4 Max is a model you can search for. That asymmetry is the entire finding, and it is the same asymmetry that has followed this family's name since September — a roster of tier names that people are already typing into an address bar with nothing behind them yet.

What makes the pairing worth a page anyway is the second-order point: the Qwen flagship that does exist, Qwen3.8 Max, is already the cheaper model per token and already the more expensive one per finished job, and that pattern is the thing to re-test rather than the tier name to wait on. If the 4 series ships with the same appetite for output tokens that the 3.8 generation shows, a lower price per million will not automatically mean a lower bill — and a reader who only follows the rate card will be surprised twice, once when the launch lands and once when the invoice does.