Alibaba released Qwen3.8-Max-Preview on July 19, 2026, and made a claim you cannot ignore: a 2.4-trillion-parameter flagship that is, in its own words, "second only to Fable 5." What it did not release is any benchmark table backing that up — no scores, no active-parameter count, no MoE configuration. The weights are promised "soon," with no date. Early independent testing runs from lukewarm to openly hostile. This page sorted the official from the leaked from the tested through the preview weeks — and on August 3, the real numbers landed: an official release with a full benchmark slate, live $2/$6 API pricing, and a dated open-weights commitment. The answer sheet is below, scored against what this page recorded.
August 3: the answer sheet
Two weeks of tracked claims, now scored against the official release (full breakdown and pricing analysis in the launch article):
| What we tracked | Where it landed |
|---|---|
| 2.4T parameters, no active count disclosed | Confirmed — with ~95B active per token finally on the record |
| Context length: 200K reported vs 1M in the UI | 1M official: 991K max input, 131K max output |
| Zero published benchmarks | Full vendor slate: Terminal-Bench 2.1 86.6 (above Opus 4.8's 84.6), SWE-bench Pro 67.7, PaperBench 93.0 — plus crowd-judged Arena entries at WebDev #4, Text #5, Vision #2 |
| "Second only to Fable 5" | Not on coding — SWE-bench Pro 67.7 vs Fable 5's 80.0. Defensible on agent and research slates, as a claim about vendor benchmarks |
| Open weights "soon," with the Qwen3.6-Max precedent | A specific, dated commitment: the 2.4T Max and a 27B variant, on Hugging Face and ModelScope, within a week of release |
| Access via Token Plan bundles only | Public API live at $2 input / $6 output, $0.25 cached |
The preview-week skepticism aged well against the marketing claim — "second only to Fable 5" did not survive contact with Alibaba's own coding numbers — and it now moves to the clock on the weights: "next week" finally has a date range, and this page logs the delivery or the slip.
Official, leaked, tested — the preview record
Official (Alibaba's own statements):
- Qwen3.8-Max-Preview is live, first through Alibaba's Token Plan subscriptions and its Qoder coding tools.
- Headline size: 2.4 trillion parameters. No active-parameter count, no expert configuration disclosed — 2.4T is a marketing number until a model card says how much of it fires per token.
- The company's positioning, verbatim from its announcement: one of the most powerful models available, "second only to Fable 5."
- Open weights are promised "soon." No date, no license named. Note the pattern: Qwen3.6-Max made a similar promise and stayed closed, hosted only on Alibaba's own platforms.
Leaked (community-circulated internal numbers, unverified):
- Coding: 12.6 points behind Claude Fable 5, with a claimed 73.2% of coding outputs matching Fable 5's.
- Cowork, a 400-task agent evaluation: ahead of Opus 4.8 Max by 5 points, Kimi K3 by 8, GLM-5.2 by 17.1, and its own predecessor Qwen3.7-Max by 44.4.
- Treat every one of these as a leak. Alibaba has published none of them, and leaked internal evals have a way of shrinking on contact with independent testing.
Tested (early hands-on, first 24 hours):
- The harshest public verdict came from an independent developer benchmark: "Not Fable 5 level. Not even remotely close" — with the marketing called misleading outright.
- Chinese-community testing (where the model has been hammered hardest) reports a stiff default style that improves with very explicit prompts, occasional identity confusion, and rendering-pipeline output that testers ranked below GLM-5.2 until prompted in detail.
- Context length is itself unsettled: 200K is the commonly reported figure, while Token Plan's own UI offers settings up to 1M. Until a model card exists, both are provisional.
- *Week-one shift (July 24):* the verdict is drifting from hostile toward mixed-positive. In one widely-shared test, a backend project built with Qwen3.8-Max passed a Claude Fable 5 code review with "high code quality" praise — though the long "thunder-thinking" pauses and occasional domain misreads persist. Alibaba also ran a weekend free-unlimited promo on the smaller Qwen3.6-27B: the distribution push is in full swing.
The honest summary: the claim is frontier-adjacent, the evidence is not yet. That does not mean the model is weak — Qwen3.7 was genuinely good, and launch-day harness problems routinely make strong models look bad. It means nobody outside Alibaba can currently tell.
The pattern this fits
Qwen3.8 is July 2026's third trillion-class Chinese release in three weeks — after Kimi K3 (2.8T, July 16) and alongside the imminent DeepSeek V4 stable — with Grok 4.6's 2T training reportedly finishing within days. The arms race is real, and so is a newer, less charming pattern: the open-weight promise as a launch accessory. K3 shipped with weights committed "by July 27." Qwen3.8 ships with weights "soon," from a lab whose previous Max release quietly never opened. Announcing openness now buys the open-source halo; delivering it later (or not) is a separate decision. We track the difference on this page, because for developers it is the entire difference between a model you can audit and a pricing page.
What you can actually use today
The staging delay resolved with the official release: Qwen3.8-Max's public API is live at $2/$6 (cache $0.25) — no Token Plan bundle required. What is available now behind one OpenAI-compatible key on Turiloop: Qwen3.7-Max and Qwen3.7-Plus — next to GLM-5.2, DeepSeek V4 and Kimi. International card, no Chinese phone number. Qwen3.8 joins the lineup as its listing rolls out — switching will be a model-id change — and the pricing calculator picks it up the same day.
FAQ
What is Qwen3.8? Alibaba's flagship model family: previewed July 19, 2026 as Qwen3.8-Max-Preview, released officially on August 3 — 2.4 trillion parameters, ~95B active, 1M context, $2/$6 API pricing, with open weights committed for the week after release. Full analysis in the launch article.
Is Qwen3.8 open source? Not yet — but the commitment became specific at the August 3 release: weights for both the 2.4T Max and a 27B variant, on Hugging Face and ModelScope, promised within a week. It would be the first Max-class Qwen ever opened; license terms are still unnamed. This page updates on delivery.
Is Qwen3.8 really second only to Fable 5? The official numbers now split the verdict. On coding, no: Alibaba's own SWE-bench Pro figure is 67.7 against Fable 5's 80.0. On agent and research slates (Terminal-Bench 2.1 86.6, PaperBench 93.0) the claim is defensible — as a statement about vendor benchmarks awaiting independent reruns.
How do I access Qwen3.8? The official API is live at $2/$6 since August 3. Through OpenAI-compatible gateways like Turiloop, Qwen3.7-Max and 3.7-Plus work today with an international card and no Chinese phone number; Qwen3.8 joins as its listing rolls out.
How does Qwen3.8 compare to Kimi K3? With real numbers now on both sides: K3 keeps Terminal-Bench 2.1 (88.3 vs 86.6) and a month of independent scrutiny; Qwen3.8-Max counters on research and OS-control slates at $2/$6 against K3's $3/$15 — under half the output price. See the K3 breakdown and the Qwen3.8 launch article.