Qwen 3.8 Max Preview (2.4T): a frontier claim shipped without a single benchmark

On Sunday 19 July 2026, at the World AI Conference in Shanghai, Alibaba's Qwen team previewed Qwen3.8-Max-Preview and said it is second only to Claude Fable 5. It shipped that sentence with no model card, no benchmark table, no license, no activated-parameter count and no first-party per-token price. Most write-ups have reprinted the claim and stopped at no benchmarks yet. This one measures the claim instead: we pulled the predecessor's independent score and worked out exactly how large a jump Alibaba is silently asserting.

What Alibaba actually announced

On Sunday 19 July 2026, during the World AI Conference in Shanghai, Alibaba's Qwen team previewed Qwen3.8-Max-Preview. The announcement was a social post, published in identical wording from the @Alibaba_Qwen and @alibaba_cloud accounts, reported by SCMP, the-decoder and MarkTechPost the same day.

The post says the model is launching and going open-weight soon, gives a figure of 2.4 trillion parameters, says the model is continuously evolving, and claims it is comparable to leading frontier models and second only to Fable 5. (The original writes compatible where comparable is clearly meant.) That is the entire technical disclosure. There is no attached model card, no benchmark table, no evaluation methodology, no license file and no repository.

Here is what is on the record, with its provenance labelled:

• 2.4 trillion total parameters — vendor-reported, unverified, no model card to check it against.
• Sparse Mixture-of-Experts — inferred, not announced. MarkTechPost explicitly flags the architecture as an inference from Qwen's history. Alibaba has not confirmed it and has published no activated-parameter-per-token figure, which is the number that actually determines serving cost and speed.
• Multimodal input — text, images, video and documents; described as the first Qwen model above 1 trillion parameters to accept multimodal input. Vendor-reported.
• Context window — approximately 1 million tokens. A precise 983,616-token figure with 131,072 max output circulates from integration metadata, but it appears on neither help.aliyun.com nor docs.qwencloud.com, both of which omit context specifications entirely. Treat the exact number as unconfirmed.
• Capability direction — Alibaba says it should beat Qwen3.7-Max especially at coding, full-stack development, data analysis and office workflows. Vendor-reported, with no supporting numbers.

And the part that matters most: as of 21 July 2026, Qwen3.8-Max does not appear anywhere on the Artificial Analysis Intelligence Index leaderboard. We fetched it directly. The board lists 186 models. Qwen entries stop at Qwen3.7 Max, Qwen3.6 Plus and the Qwen3.5 variants. No independent evaluation of Qwen3.8 exists yet, from anyone.

Announced vs evidenced: the audit

The most useful thing we can do with an announcement like this is line it up attribute by attribute against what a buyer can independently verify. One row per attribute, one column for the claim, one column for the evidence, one column for where you can check it yourself.

Nine attributes. Nine cells in the evidence column that are empty, inferred or contradicted. That is not a normal condition for a model being publicly ranked second in the world, and it is why we are not writing a review — there is nothing to review. If you want a model page where the numbers actually exist, our Qwen3-Max review covers the predecessor line that has real published scores and a real rate card.

How big the claim actually is

Here is the number nobody else pulled. Qwen3.7-Max — the direct predecessor — scores 46 on the Artificial Analysis Intelligence Index, ranked #27 of 186, priced at $2.50 per million input and $7.50 per million output with a 1M context. We fetched that page directly on 21 July 2026.

Now put that next to the top of the same board on the same day: Claude Fable 5 (with fallback) 60, GPT-5.6 Sol (max) 59, GPT-5.6 Sol (xhigh) 58, Kimi K3 57, GPT-5.6 Sol (high) 56, Claude Opus 4.8 (max) 56.

For second only to Fable 5 to be true on this index, Qwen3.8-Max has to score above Kimi K3's 57 and below Fable 5's 60. Strictly, to be second it also has to clear GPT-5.6 Sol at 59. That is a single-generation jump of roughly +12 to +13 points, and a move from rank #27 to rank #2.

That is the information gain. Every other outlet reprinted the sentence. The sentence is not vague — it is a precise, arithmetic, falsifiable assertion about a public leaderboard, and it is a very large one.

Two things are worth saying plainly about that dashed band. First, it is not a prediction — we are not asserting Qwen3.8 will or will not land there. Second, a +12-point generational jump is not impossible; it is just extraordinary, and extraordinary jumps are exactly the ones that get published with evaluation harnesses attached. This one wasn't.

The pricing problem: Credits, not tokens

You will read that Qwen3.8-Max-Preview has no pricing. That is slightly too strong, and the accurate version is more interesting.

Alibaba has published no pay-as-you-go per-million-token rate card for it, and the model is absent from Alibaba Cloud's international Token Plan model table (which stops at qwen3.7-max, qwen3.7-plus, qwen3.6-plus and qwen3.6-flash, alongside DeepSeek, Moonshot, Zhipu and MiniMax models). But it is listed on the Model Studio / Qwen Cloud Personal Token Plan, with real subscription tiers:

• Lite — 60 CNY/month, limited-time 39, 2,500 Credits per 7 days
• Standard — 180 CNY/month, limited-time 139, 10,000 Credits per 7 days
• Pro — 600 CNY/month, limited-time 499, 40,000 Credits per 7 days

Note the unit: Credits, not tokens. The documentation adds that during the promotional period, Credits consumption is as low as 10% of the standard rate, effectively providing 10x the usage, with a further off-peak discount for calls between 22:00 and 08:00 UTC+8. (The page's compounded arithmetic does not reconcile, so we are not quoting a combined percentage.) Access also runs through the Qoder and QoderWork agentic platforms.

Third-party gateways have already begun listing per-token rates for qwen3.8-max-preview, and they do not agree with each other. NanoGPT lists $1.50 per million input and $5.00 per million output (plus $0.15 cache read); AIHubMix lists $0.17 and $0.51, roughly a tenth of NanoGPT's figures — consistent with the 10% preview discount being baked into one listing and not the other, though neither vendor says so outright. That is a roughly 9x spread between two resellers of the same preview model, with no first-party rate card to reconcile them against. The disagreement is the finding: nobody outside Alibaba knows the standard rate, so nobody can tell you what the discount is a discount from.

So the honest framing for anyone doing capacity planning: access is metered in opaque subscription Credits, at a temporary discount off a standard rate Alibaba has never published for this model, and the only per-token numbers in circulation are reseller listings that differ by an order of magnitude. You cannot derive a defensible cost per million tokens. You therefore cannot derive a cost per task, cannot compare it to anything in your stack, and cannot put a line item in a budget. For an engineering team, that is a harder blocker than the missing benchmarks — a quality claim you can eventually test, but a unit price that does not exist cannot be modelled at all. For comparison, our Qwen API pricing breakdown shows what the published Qwen tiers actually cost per token.

For context on the international side, Alibaba's Team Edition Token Plan runs USD 30/seat/month for 25,000 Credits (Standard), USD 100 for 100,000 (Pro) and USD 200 for 250,000 (Max) — but qwen3.8-max-preview is not on that plan's model list.

One aside worth noticing: search results for this model are already flooded with pages titled things like Qwen 3.8 Max REVIEW: I tested it and Qwen 3.8 benchmarks, published for a model with no public benchmarks and no general API. Hands-on reviews of software nobody outside a preview can access are a useful reminder to check whether a page's numbers have a source you can click.

What a well-evidenced launch looks like

Three days before WAIC, on 16 July 2026, Moonshot AI released Kimi K3 — a 2.8-trillion-parameter MoE with a 1M-token context, priced at $3.00 per million input and $15.00 per million output. Within days it was independently measured: 57 on the Artificial Analysis Intelligence Index, rank #4 of 186, with Artificial Analysis also noting the trade-offs (slow at 36.9 tokens/sec, and very verbose). It was also reported as #1 on the blind Frontend Code Arena, ahead of Claude Fable 5 — we could not verify that leaderboard directly, so treat it as reported by Tom's Hardware and other outlets rather than as first-hand.

One correction to how K3 is often described: it was not an open-weight launch. It shipped API-only on 16 July; weights are dated 27 July 2026 under an expected Modified MIT license, and as of 21 July the newest model on Moonshot's Hugging Face organisation is still K2.7 Code. Open-weights-dated, not open-weight. We cover it in full in our Kimi K3 review.

This is not a story about which country's labs overclaim. K3 is from a Chinese lab too, launched three days earlier, and it is the example of doing this right. The variable is evidence standards, not origin.

An open-weight Max release would genuinely be a departure — Alibaba has historically open-sourced the smaller and mid-tier Qwen models while keeping the flagship Max line proprietary and API-only, which is exactly the state Qwen3.7-Max is in today. If it happens, it changes the open-weight landscape materially; see our running best open-source LLM in 2026 comparison and the Qwen vs DeepSeek head-to-head for where the open tiers currently sit.

What would settle it

The nice thing about a numeric claim is that it is falsifiable. Two events would resolve this within days of general access:

1. An Artificial Analysis Intelligence Index entry above 57. That clears Kimi K3. Above 59 clears GPT-5.6 Sol and makes the literal claim of second place true. Below 57 does not.
2. A blind arena placement above Claude Fable 5. Human preference under blind conditions is the check that vendor-selected benchmarks cannot game.

Two dates to watch: 27 July 2026, the date on Kimi K3's promised weights, and the Qwen3.8 open-weight release, which currently has no date at all. A third signal, easy to miss: the day qwen3.8-max-preview appears on the international Token Plan model table with a per-token rate, it becomes a product you can plan around rather than a preview you can only sample.

A decision rule for teams

If you are deciding what to do this week, the rule is short:

• Do not migrate anything. There is no first-party unit price, no general availability and no independent quality signal. Migration decisions need at least two of those three.
• Do not rewrite your model ranking on the strength of a launch post. Rank on measurements you can reproduce — either a public independent index or your own harness. Our best LLM in 2026 roundup is built on the former, and it will get a Qwen3.8 row the day one exists.
• Do prepare the test. Write down now what score Qwen3.8 has to hit on your tasks to be worth switching cost. If your current model is doing the job at $2.50/$7.50 per million, a 2.4T model needs to beat it by enough to cover the migration, not just to top a leaderboard.
• Do watch the weights. An open-weight 2.4T Max model matters even at a lower quality tier, because self-hosting and fine-tuning change the economics in a way an API score does not capture.

None of this is a prediction that Qwen3.8 will disappoint. Qwen has shipped genuinely strong models, and the coding and agentic direction Alibaba describes is plausible. The point is narrower and, we think, more useful: as of 21 July 2026, the correct amount of confidence to have in the second-only-to-Fable-5 claim is zero — not because it is implausible, but because nothing has been published that could make it more than zero. Come back when there is a number.

FAQ