
Qwen4-Max vs Qwen4 27B: A 4.5-Fold Price Gap and a Model You Cannot Download
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 96 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0717 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 87 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 72 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 130 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 65 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 43 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 193 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 152 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 119 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 219 tok/s
The interesting question about Qwen4-Max is not whether it is good. It is whether you should be paying $1.50 and $4.50 per million tokens for a model that thinks in 262,144-token gulps when a 27-billion-parameter sibling — Qwen4 27B — is supposed to do most of the same work on hardware you own. The catch is that Qwen4 27B does not exist yet as anything you can run. Alibaba showed it on stage at the Apsara Conference on 22 September 2026 alongside Qwen 4 Max, Qwen 4 Flash and Qwen 4 Plus, and has published no model card, no weights, no context length, no price and no benchmark for it since. Qwen4-Max, by contrast, went metered on 16 November with a 1,000,000-token window and an Artificial Analysis Intelligence Index of 43. So the useful version of this comparison sets the flagship you can call against the small model you cannot, and answers the question that actually decides budgets: what does the 4.5-fold token-price difference buy, and what would the 27B have to do to earn its place.
Two price cards, one of which has numbers on it
Start with what is written down. Everything in this section is a published figure; nothing is inferred.
• Qwen4-Max, Singapore — $1.50 per million input tokens, $4.50 output, flat across the full 1,000,000-token window. Cache reads $0.075 per million, five-minute writes $1.875, one-hour writes $3.00.
• Qwen4-Max, other regions — $1.413 and $4.24 across Frankfurt, Virginia, Tokyo, Hong Kong and Beijing, a 5.8% discount on the Singapore card for identical limits.
• Qwen4-Max, independent — AA Index 43 at $3.27 of evaluator-billed cost per index task, blended $0.80 per million, 127M output tokens across the run.
• Qwen4 27B — no price. Not "unannounced"; specifically, no row exists anywhere because no model exists to attach one to.
The nearest thing to a small Qwen you can run today is Qwen3.8-27B, the Apache-2.0 dense multimodal model from the previous generation. Alibaba lists it at $0.50 and $3.00 per million on its own API; we self-host the weights and charge $0.33 and $2.40, because open weights have no per-token vendor cost to pass through. Artificial Analysis scores it 20 on the same index — twenty-three points below Qwen4-Max, which is the whole argument in one number.
So the honest ratio is not measured against a phantom. It is 4.5-fold on input and 1.9-fold on output when you compare Qwen4-Max with the last 27B that actually shipped — and twenty-three index points of capability for the difference. Whether that is a bargain depends entirely on what your workload spends its tokens on, which is a question about your traces rather than about either vendor.

What a 27B buys that a 2.4-trillion-parameter flagship cannot
Three things, and none of them is quality.
The first is a fixed bill. A hosted flagship is a variable cost that scales with traffic and with how verbose the model feels that day; an open 27B on your own hardware is a capital cost and a depreciation schedule. At $0.33 and $2.40 per million, a team spending $3,000 a month on the flagship is spending roughly $700 on the small model for the same token volume — and the moment that volume doubles, one of those numbers moves and the other does not.
The second is latency you control. A 27B dense model on a single node answers in a shape you can predict, because there is no queue behind it and no long-horizon thinking budget you did not ask for. Qwen4-Max defaults to reasoning_effort: high with a 131,072-token thinking budget, and a request that sets nothing has reserved more output headroom than most jobs will use. That default is right for a flagship and wrong for a classifier, a router, or the thousand small decisions an agent makes between the interesting ones.
The third is the licence. Apache-2.0 weights mean the model can be quantised, fine-tuned, embedded in a product you ship, and run in a jurisdiction where the API is not available. Qwen4-Max is proprietary, served from six regions, and returns text only — no weights, no fine-tuning, no on-premise deployment. Alibaba's documentation says fine-tuning and model deployment are not supported for it anywhere.
What a 27B cannot buy you is the thing Qwen4-Max was built for. The 2.4-trillion-parameter backbone with 71 billion active parameters, the 328 billion N-gram embedding parameters, the Gated DeltaNet layers compressing history into fixed state for each Qwen Sparse Attention layer — none of that architecture transfers down to a 27B dense model, and neither does the 1,000,000-token window. The vendor's own framing of the family is that smaller models are "tuned for what small models do best," which is a graceful way of saying they are not the same product.
The one thing Alibaba has said about Qwen4 27B, and what it is worth
In the Qwen4-Max release post on 16 November, the closing paragraph reads: "And Max is only the first look at Qwen4. Smaller models in the family are coming to open source, built on the same lineage and trained with the same generation of techniques, tuned for what small models do best."
That is the freshest and most specific statement in existence about the 27B tier, and it is worth exactly what it says — which is less than it appears to. "Coming to open source" is a commitment to a licence, not a date. "Same lineage" is a claim about architecture descent, not about a released checkpoint. And the sentence does not name 27B at all; the tier name comes from the September stage announcement, where Qwen's LLM lead showed four models and said only that the family would arrive soon. Between the two statements there is a two-month gap with four tiers promised, one delivered, and the open-weight one — historically the tier that ships first under downloadable weights — still absent.
What would make this page out of date is checkable in three places, and all three are the ones that have signalled every Qwen release. A repository in the Qwen organisation on Hugging Face with a Qwen4 name; the organisation's newest model repositories today are all Qwen3.8, last updated in early November. A model card on qwen.ai naming the tier with a context length and a licence. Or a pricing row for a hosted 27B, which would tell you what Alibaba thinks the tier is worth before the weights land. Until one of those moves, any parameter count, context window or score you read for Qwen4 27B is extrapolation from Qwen3.8-27B and should be labelled as such.
Worked example: the same agent loop on both cards
Take a support-automation agent: 40,000-token system prompt and tool schema, a 6,000-token conversation, eight tool calls per ticket, 900 output tokens per call, no thinking mode, 20,000 tickets a month. Cache the prefix. This is arithmetic on published list prices.
• Qwen4-Max, Singapore, uncached — 46,000 input tokens per call at $1.50 per million is $0.069; eight calls is $0.552 per ticket. Output at 900 tokens and $4.50 per million is $0.00405 per call, $0.0324 per ticket. $0.584 per ticket, $11,690 a month.
• Qwen4-Max with the prefix cached — the 40,000-token prefix is written once at $1.875 per million and read on the next seven calls at $0.075. Per ticket the prefix costs $0.075 for the write plus $0.021 for seven reads; the moving 6,000 tokens cost $0.072 across eight calls. Input lands at $0.168, output stays $0.0324. $0.20 per ticket, $4,010 a month — and the cache did more for the bill than any model swap would.
• Qwen3.8-27B at our self-hosted rate — 46,000 input tokens per call at $0.33 per million is $0.0152; eight calls $0.121. Output at 900 tokens and $2.40 per million is $0.00216 a call, $0.0173 per ticket. $0.139 per ticket, $2,780 a month.
Two things fall out. Caching the prefix on the flagship saves 66% and brings it within 1.4× of the small model's uncached price — the gap between tiers is smaller than the gap between using the cache and not using it. And the small model's advantage is real but not transformational at this token volume: about $1,200 a month on 20,000 tickets. Where it becomes transformational is at ten times that volume, or where the work is simple enough that a 20-index model is sufficient, which for tool-call routing and classification it usually is.
Where the flagship is not optional
There is a category of work in which the 27B tier is not a cheaper substitute but a non-starter, and being clear about it is the difference between a useful comparison and a sales page.
• Long-context work above 256,000 tokens — Qwen3.8-27B stops at 262,144; Qwen4-Max carries 1,000,000, with a 991,808-token input ceiling in non-thinking mode. A 400,000-token contract review does not fit in the small model at all.
• Long-horizon agent runs — Alibaba's agentic claims for Qwen4-Max are a 48-hour autonomous materials-discovery run on a wet-lab platform and a legal due-diligence pass over 198 working papers producing an 86-page report. Whatever one thinks of vendor demonstrations, they describe a class of task that a 20-index model does not attempt.
• Video input at length — Qwen4-Max takes video and a 1M window; the 27B family takes video at 262K. The arithmetic on frames per second of footage is unforgiving.
• Secure coding — Alibaba reports Last Secure Code Bench function-level pass rates rising from 22.0% to 31.8% between generations on the Max tier, and code that is secure at function, file and repository granularity simultaneously from 8% to 14.7%. Those are vendor figures and unreproduced, but they are the kind of number a security review asks for and a small model's card does not carry.
The pattern is that the flagship earns its price on the tasks that are long, multimodal, or expensive to get wrong, and not on the tasks that are numerous. Most production systems contain both kinds, which is why the answer to "which one" is usually "both, routed."
Running the mixed state on one key
OrcaRouter carries the Qwen3.8 family — Qwen3.8-Max, Qwen3.8-Flash and Qwen3.8-27B — on one OpenAI-compatible endpoint at provider list price with 0% markup, and the 27B is the one we run on our own infrastructure from the Apache-2.0 weights, which is why it prices at $0.33 and $2.40 rather than Alibaba's $0.50 and $3.00. It does not carry Qwen4-Max: that model is reachable through Alibaba's own Model Studio API and several third-party platforms, and is not in the catalogue as of 16 November.
What that makes possible today is the routing pattern this comparison argues for, without waiting for the 27B. One API for more than 200 models means the tier decision stops being a procurement decision: a request that fits in 256K and looks routine goes to qwen/qwen3.8-27b; a request that is long, multimodal, or flagged as high-stakes goes to a frontier model on the same key; the routing DSL expresses that split as a rule rather than as a branch in your application code. Automatic failover means the cheap tier can carry the volume without becoming a single point of failure when it has a bad afternoon, and a price cut from any vendor lands on your key the same day because the list price is passed through rather than marked up.
Model fusion is the third option in that toolkit and the one that maps least well onto this particular pair — a panel of models answering together costs more than either alone, and the reason to use it is disagreement on hard questions, not volume. Worth knowing it exists; not worth reaching for on a support-ticket router.

What to do before the weights land
If your reason for waiting on Qwen4 27B is that you want a small open model you can run yourself, the wait is not costing you much: Qwen3.8-27B is Apache-2.0, carries a 256K window, reads text, images and video, and posts 90.3 on LiveCodeBench v6 and 89.2 on GPQA Diamond on Qwen's own card — vendor figures, and strong ones for the size. Start there, measure what a 20-index model does to your task success rate, and you will know before the Qwen4 checkpoint arrives whether the tier is right for your workload at all.
If your reason is cost, cache the prefix first. The worked example above shows a 66% saving on the flagship from one request parameter, which is larger than the gap between the flagship and the small model on the same uncached basis. Teams routinely shop for a cheaper model when the expensive part of their bill is a prefix they are re-reading at full price.
And if your reason is capability, no 27B will substitute. The twenty-three index points between Qwen3.8-27B and Qwen4-Max are the difference between a model that handles routine work well and one that attempts the long, multimodal, high-stakes tasks — and Alibaba's own roadmap language says the small models are being tuned for what small models do best, which is a promise of efficiency rather than of parity. Build the routing split now, on the tiers that exist, and the day a Qwen4 27B lands on Hugging Face it is a catalogue entry and a rule change rather than a migration.

