Mercury 3 vs Qwen 4 Max: One of These Models Does Not Exist Yet

Mercury 3 has been callable since October 19, 2026: a closed diffusion model from Inception with a 1M-token context, 131,072 output tokens and a $0.20/$0.70 rate card. Q​wen 4 Max has been callable for zero days, because it has not been released. A​libaba showed the name on stage at its Apsara Conference in Hangzhou on September 22, 2026, alongside Q​wen 4 Flash, Q​wen 4 Plus and Q​wen 4 27B, described the family as training on a new-generation architecture, and said it would arrive soon. As of October 20 there is no model card, no API identifier, no context window, no price, no licence and no benchmark score for any Q​wen 4 tier. So this is not a model-versus-model article in the usual sense. It is an article about what you can actually buy today, what a stage announcement is worth, and why the real alternative to Mercury 3 on the Q​wen side is Qwen3.8-Max rather than the model in the headline. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

What is actually being compared

Two different kinds of object. Mercury 3 is a shipped product with a published rate card, a documented operating envelope and a vendor launch post that makes specific numeric claims. Q​wen 4 Max is a tier name attached to a roadmap. Every fact in the Q​wen column below is a fact about the announcement, not about the model:

• Release date — Mercury 3 October 19, 2026 vs Q​wen 4 Max none announced; the line was described as "in training"

• Context window — 1,000,000 tokens vs unpublished

• Maximum output — 131,072 tokens vs unpublished

• Price — $0.20 per 1M input and $0.70 per 1M output, with a 50% launch discount vs unpublished, in any currency

• Modality — text only vs unpublished; the Apsara framing stressed agentic and tool-using work

• Weights — closed vs unpublished; the 27B tier in the same family is the one historically shipped open

• Throughput — 1,068 tokens per second, vendor-reported vs unpublished

• Independent benchmark score — none yet for Mercury 3 vs none possible for Q​wen 4 Max

• Parameter count — undisclosed, "more than 40× Mercury 1" vs unpublished; the 5-to-10-trillion figure in circulation belongs to the Q​wen 4.5 and Q​wen 5 roadmap, not to Q​wen 4 Max

• What you can do with it today — call it through Inception's API vs nothing

That last line is the comparison. Nine of the ten rows above are "unpublished", and a spec table that is nine-tenths blank is not a spec table, it is a press release with column formatting.

What a conference announcement is worth

A​libaba's Q​wen line has a track record worth taking seriously, which is why the announcement got coverage at all. The 3.8 generation shipped a 2.4-trillion-parameter flagship at $2.00/$6.00, published weights for its 27B tier, and released sub-versions on a monthly cadence — Qwen3.8-Max went GA on August 3, 2026, with a (0902) refresh on September 2, a 27B open-weight release on August 14 and a Flash-Next model on August 26. That is a lab that ships, on a schedule, with weights.

It is also a lab that announces ahead of shipping. The Q​wen 4 lineup was presented on September 22 with no dates attached to any of the four tiers, and the four weeks since have produced no model card, no repository, no Hugging Face checkpoint and no entry on the independent boards. The realistic planning assumption is not "Q​wen 4 Max arrives next month" but "Q​wen 4 Max arrives when a model card appears", which is an unresolvable date and therefore not a date you can build a migration around.

What a roadmap announcement is genuinely good for is a different kind of decision. If you are choosing between vendors today, the announcement tells you A​libaba intends to keep competing at the flagship tier, which is a reason to stay on its current generation rather than leave the family. If you are choosing a model to deploy this quarter, it tells you nothing you can act on.

The 5-trillion-to-10-trillion-parameter number that travelled alongside Q​wen 4 coverage deserves a specific correction, because it will keep propagating. It is a forward-looking statement about the generations after Q​wen 4 — the Q​wen 4.5 and Q​wen 5 timeframe — and A​libaba has not attached any parameter count to Q​wen 4 Max. A Q​wen 5-class size claim printed next to a Q​wen 4 model name is a category error with a long half-life.

The Q​wen you can actually call

If the question behind the search is "should I use Mercury 3 or A​libaba's flagship", the A​libaba flagship is Qwen3.8-Max and it has a real record. The comparison that matters is therefore Mercury 3 against the model that exists:

• Release — Mercury 3 October 19, 2026 vs Qwen3.8-Max August 3, 2026, refreshed September 2

• Context — 1,000,000 tokens vs 983,616 tokens as recorded by Artificial Analysis; the vendor quotes a flat 1M

• Price — $0.20/$0.70 at list, $0.10/$0.35 during launch vs $2.00/$6.00, with cached input at $0.25

• Cached input — $0.02 ($0.01 at launch) vs $0.25; Mercury 3 is an order of magnitude cheaper on cache reads

• Independent Intelligence Index — no entry vs 45.4

• Cost per index task — not measured vs $5.41

• Terminal-Bench 2.1 — not measured vs 0.8876, the highest in this batch

• AA-LCR long-context reasoning — not measured vs 0.8033

• GDPval-AA — not measured vs Elo 1665.58

• AA-Omniscience — not measured vs 12.0

• Weights — closed vs closed for the Max tier; the 27B tier of the same generation is open

The shape of that comparison is familiar from the rest of this batch and worth stating without hedging: Mercury 3 is roughly a tenth of Qwen3.8-Max's price on input and a twelfth on output, with a marginally larger context window and no independent score. Qwen3.8-Max is the measured model, and its measurements are good where they exist — 0.8876 on Terminal-Bench 2.1 is the strongest reading in this five-article set — but its $5.41 cost per completed index task is the highest too, and it is the number that decides whether the capability is affordable at volume.

That is the actual decision a reader of this page faces today. Not Mercury 3 versus a Q​wen 4 model that does not exist, but Mercury 3 versus the Q​wen model that A​libaba will sell you, with the Q​wen 4 announcement as a reason to watch the family rather than a reason to wait for it.

How to hold a place for Q​wen 4 Max without betting on it

The reasonable position for a team that expects Q​wen 4 Max to matter is to keep the option cheap. Three things make that possible, and none of them require waiting.

First, write model selection as configuration rather than code. Mercury 3 and Qwen3.8-Max are both reached through OpenAI-compatible APIs, so a stack that resolves the model identifier at request time can adopt a Q​wen 4 tier on the day a model card appears without a deploy. This is the same discipline that makes a vendor's deprecation notice a config change instead of an incident, and Inception has just demonstrated how fast those notices arrive: Mercury 2 and Mercury 2.5 close to new users as of October 19, with existing access ending November 15.

Second, benchmark the workload, not the model. The vendor claims for Mercury 3 are a 50-task Mind2Web result in its own harness and testimonial quotes from four partners. Those tell you the model is fast in production-shaped work; they do not tell you how it handles your documents. Thirty of your own eval cases, run against Mercury 3 and Qwen3.8-Max on the same prompts, will out-inform every launch table published this month.

Third, keep the measured model on the path that cannot fail. OrcaRouter carries qwen/qwen3.8-max at A​libaba's list price with 0% markup and automatic failover across 200-plus models on one key, so the incumbent side of the comparison is a routing table rather than a new contract. Mercury 3 is not on the router — it is served through Inception's own API and several third-party platforms — which means an honest trial puts it on a slice of traffic behind the vendor's API while the measured model keeps the production path. If Q​wen 4 Max ships into the router later, it slots into the same policy without a second integration.

The question to ask before the next announcement

Every year produces a handful of model names that get searched for months before a model exists behind them, and the cost of that gap is paid by teams who plan against the announcement. The useful habit is a one-line test: can I call it, and is there a price? Mercury 3 passes both today. Q​wen 4 Max passes neither, and will not until A​libaba publishes a model card.

What to watch, specifically: the Q​wen 4 27B checkpoint rather than the Max tier's benchmark table, because the 27B is the tier that historically ships open and its licence and size will tell you how serious the open-weight commitment is this generation. Then the Max tier's rate card, because a Q​wen flagship that undercuts $2.00/$6.00 would change the affordability argument in this comparison rather than the capability one. And on the Mercury side, a first independent score — the family's only measured member, Mercury 2.5, sits at 12.3 on the Intelligence Index, so the flagship's quality envelope is entirely a vendor claim until the board weighs in.

Until then, the comparison is between a model you can rent this afternoon at a price that makes per-request economics trivial and a model that exists in a conference hall in Hangzhou. One of those is a decision. The other is a calendar entry.