A generated 16:9 title card reading 'Qwen4-Max vs Kimi K4', subtitled 'One shipped this week; the other has never shipped at all', with two rounded cards: a documented GA model with six regions and pay-as-you-go pricing, and a name with no model card, no API identifier and no announcement. OrcaRouter logo bottom-right.
Guides & Insights

Qwen4-Max vs Kimi K4: One Shipped This Week, the Other Has Never Shipped at All

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Search for K​imi K4 and you will find a great deal of writing about a model that does not exist. Search for Qwen4-Max and, as of 16 November 2026, you will find a model you can call with a published rate card, six regions and a documentation page timestamped 09:46 that morning. Those are not comparable states of affairs, and treating them as two sides of a matchup is the single most common mistake in this corner of the market. So this page does two jobs: it establishes what K​imi K4 is — a name with no model card, no API identifier, no weights, no price and no announcement behind it — and then it runs the comparison a reader can actually act on, between Qwen4-Max and Kimi K3.1 Pro, the model Moonshot AI ships today and the highest-scoring open-weights model on the Artificial Analysis Intelligence Index.

That second comparison is more interesting than the one you typed in, because it is the one place where Qwen4-Max meets a model that beats it on measured capability while losing to it on price. It is also the comparison that shows why the missing Qwen4-Max benchmark is expensive: against Kimi K3.1 Pro there is a real number on one side of the table and a blank on the other.

A generated two-column scoreboard titled 'Qwen4-Max vs Kimi K3.1 Pro - the scoreboard', comparing $1.50 and $4.50 per million tokens against $2.00 and $10.00, a 1,000,000-token context against 1,048,576, a 131,072 maximum output against a 1,000,000-token ceiling, no independent score against 46.5, and no weights against an open parent under a custom licence, over a footer reading 'Qwen4-Max per Alibaba; Kimi figures per Moonshot and Artificial Analysis Index v5.0.1.' OrcaRouter logo bottom-right.

The K​imi K4 check, done in the open

Claims that something has not shipped decay, so here is the evidence rather than the verdict. Moonshot AI's developer pricing documentation for the K​imi API enumerates kimi-k3.1-pro and kimi-k3 as the K3-series models, with a 1,048,576-token context window, plus the K2 series, which the same page says will be fully retired by the end of December 2026. No kimi-k4 identifier appears. The platform's model list does not carry one. The lab's own blog index carries no K4 post. A repository search under the moonshotai organisation returns nothing above K3.1 Pro.

The one artefact that surfaces for the name is a third-party re-upload on Hugging Face, created in July 2026, with a handful of downloads, no documentation, and a tag set copied wholesale from Kimi K3. It is not a Moonshot artefact and this page does not treat it as one. A shipped Moonshot flagship leaves a checkable trail within days — a model card with a parameter count, a rate-card row, an API identifier, a leaderboard entry. A name that has left none of that is not a quiet release; it is a name.

Worth saying plainly, because the framing of this page could be misread: the fact that K4 is not out says nothing about Moonshot's trajectory. The lab has shipped twice in four months, and the second of those, K3.1 Pro on 24 October 2026, is the best open-weights result on the current index. The next model is a when, not an if. It is simply not a now.

What Qwen4-Max became this week

Until 16 November, Qwen4-Max existed only as qwen4-max-preview: a Token Plan-only rental opened at 10:00 UTC+8 on 11 November, with a published retirement at 24:00 on 25 November and no per-token price anywhere. The GA model is a different identifier, qwen4-max, with pay-as-you-go rates, six regions — Singapore, Beijing, Hong Kong, Frankfurt, Virginia and Tokyo — and no expiry date.

The published specification is a 1,000,000-token context, a 131,072-token maximum output in non-thinking mode, a shared 262,144-token budget for thinking plus answer, text, image and video input, and output in text only. Thinking is enabled by default with a 131,072-token budget and reasoning effort set to high. The API speaks O​penAI Chat Completions, O​penAI Responses, A​nthropic Messages and DashScope, which means an existing A​nthropic SDK pointed at a new base URL will work without a rewrite. Fine-tuning and model deployment are documented as unsupported, and no weights are published.

A screenshot of Alibaba Cloud Model Studio's 'Model releases and updates' changelog, dated 2026-11-16 and headed qwen4-max, with the 2026-11-11 qwen4-max-preview entry visible below and the documentation timestamp at the top of the page.

Qwen4-Max against the K​imi you can actually call

Kimi K3.1 Pro was a post-training release rather than a new architecture: Moonshot describes it as Kimi K3 retrained with substantially more post-training compute, including large-scale joint reinforcement learning across hundreds of downstream domains. The context window is unchanged at 1,048,576 tokens, vision and video input are unchanged, and the output ceiling is unchanged at 1M tokens. What changed most consequentially was the price: $2.00 per million input tokens and $10.00 per million output, down from K3's $3.00 and $15.00.

• Input price — Qwen4-Max $1.50 per 1M in Singapore, $1.413 elsewhere, versus Kimi K3.1 Pro $2.00

• Output price — Qwen4-Max $4.50 Singapore, $4.24 elsewhere, versus Kimi K3.1 Pro $10.00

• Cache read — Qwen4-Max $0.075 after a $1.875 five-minute write, versus Kimi K3.1 Pro $0.20 after a $2.00 write

• Context — Qwen4-Max 1,000,000 tokens versus Kimi K3.1 Pro 1,048,576 tokens, a 4.6% edge to K​imi

• Max output — Qwen4-Max 131,072 non-thinking, versus Kimi K3.1 Pro's 1M-token ceiling, a large edge to K​imi

• Independent score — Qwen4-Max unpublished, versus Kimi K3.1 Pro 46.5 on Artificial Analysis Intelligence Index v5.0.1, the top open-weights result on the board

• Weights — Qwen4-Max none published, versus Kimi K3.1 Pro's parent K3 open under a custom licence

• Vendor benchmark — Qwen4-Max none, versus Kimi K3.1 Pro at 65.7% on Terminal-Bench 4.0, vendor-reported

The output-price row is the one that reorganises a budget. At $4.50 against $10.00, Qwen4-Max is 55% cheaper on output in Singapore and 58% cheaper in the other regions. A pipeline that generates 50 million output tokens a month pays $225 on Qwen4-Max and $500 on Kimi K3.1 Pro. That is a real gap, and it is the strongest single argument for the newer model.

The output-ceiling row is the one that can cancel it. Kimi K3.1 Pro will emit a million tokens in one response; Qwen4-Max caps at 131,072 outside thinking mode and 262,144 shared with it. If your workload includes generating long documents, full-file rewrites or extended reasoning traces in a single call, the cheaper model may not be able to finish the job at any price, and a 55% discount on a request that has to be chunked into eight calls is not a discount.

Why the blank in the score column is the story

Every model comparison this blog publishes runs into the same wall with new A​libaba flagships: the documentation is unusually precise about mechanics and completely silent about quality. Qwen4-Max's page specifies the default temperature in each mode, the top-p and top-k values, the thinking budget, the exact input and output token ceilings per mode, and the six regions with their deployment scopes. It publishes no benchmark table, vendor-reported or otherwise, and there is no Qwen4 entry on the Artificial Analysis model index — the per-model URL returns a 404.

Against Kimi K3.1 Pro that asymmetry has a specific consequence. The K​imi side has a first-party number everyone can see and an independent number on the same index vintage, and both of them sit at the top of the open-weights board. The Q​wen side has a price and a spec sheet. A buyer comparing them is not comparing two models; they are comparing a measured model to a documented one, and the only way to close that gap is to measure it themselves.

Which is cheaper to do than it sounds. A hundred prompts drawn from your own traffic, run on both, at these rates, costs a few dollars. That is the entire evaluation budget required to convert Qwen4-Max from an unknown into a known quantity for your workload — and it is less than the cost of a single week of guessing wrong on output price.

The one place the two designs diverge hardest

Moonshot's K3 architecture leans on K​imi Delta Attention, a linear-attention mechanism with a fixed-size recurrent state, across 69 of its 92 sparse layers. That design is what makes a 1M-token context affordable to serve, and it is also the reason independent evaluators discount first-party long-context claims across the industry: linear-attention hybrids tend to give back a few points on needle-in-a-haystack retrieval relative to full attention. Q​wen publishes nothing about Qwen4-Max's attention design, so there is no way to know which side of that trade it sits on.

What Q​wen does publish is a prompt-caching implementation with explicit and implicit modes and a five-minute cache write at $1.875 per million against a read at $0.075 — a 96% discount on the read. K​imi's equivalent is $2.00 written and $0.20 read, a 90% discount. Both designs reward a stable prefix and punish a prompt that changes every turn, and neither changes the arithmetic on a one-shot request.

For long-context document work — the workload both vendors are selling — the practical test is the same on either model: build a needle test on your own corpus before you trust the headline window. A model that answers correctly at 900K tokens on your data is worth more than one that answers correctly at 1M on a synthetic benchmark, and only one of those two numbers is knowable without your own harness.

Calling them side by side

OrcaRouter carries Kimi K3.1 Pro as kimi/kimi-k3.1-pro at Moonshot's list price — $2.00 in and $10.00 out per million — passed through with no markup, on the same OpenAI-compatible base URL as the rest of a 215-identifier catalogue, with both /v1/chat/completions and /v1/responses endpoints. Qwen4-Max is not in that catalogue today; it is available through A​libaba's own API, where the regional rate card above applies.

A screenshot of the OrcaRouter model page for Kimi K3.1 Pro, showing the identifier kimi/kimi-k3.1-pro, $2.00 input and $10.00 output per million tokens, the 1,048,576-token context window, the large maximum output figure, text, image and video input, and the 2026-10-24 vendor date.

The reason to keep both reachable rather than picking one is that the decision between them is workload-shaped, not model-shaped. Long-generation work wants K​imi's output ceiling. High-volume generation where responses are short wants Qwen4-Max's output price. Anything where a wrong answer is expensive wants the model with a score until you have generated a score for the other one. A routing policy expresses that as rules rather than as an architecture decision, and automatic failover keeps the newer of the two from becoming a single point of failure while you are still collecting evidence.

Where this leaves the search you started with

If you came for K​imi K4, the answer is that it has not shipped, nothing about it has been announced, and the model to evaluate today is Kimi K3.1 Pro — which, on the current index, is the strongest open-weights model available to anyone. Watch for a model card rather than a rumour: a document carrying parameters, context, price and an API identifier together is the signal this blog has used all year to tell a shipped model from a name in circulation.

If you came for the Qwen4-Max side, the answer is that it is real, it is priced, and it is cheap in exactly one dimension that matters — output tokens — while capping at a fraction of what the K​imi model will emit in a single response and carrying no measurement of quality at all. Buy it where output volume dominates and responses are short. Test it before you buy it anywhere else. And treat the gap between $4.50 and $10.00 as an invitation to run your own eval, because that is the only number in this comparison that is currently free.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily