Alibaba (Qwen)

Qwen3.8 Max

Released Jul 19, 2026 · knowledge cutoff unpublished

Status: Superseded
Location: China
Modality: Multimodal
Context window: 984K
Max output: 131K
Speed: —
Price ($/MTok in / out): $2 / $6
Cost per task: —
Open weights: No

Benchmarks4 of 8 reported

GPQA Diamond

92.7%

Terminal-Bench 2.1

81.3%

Humanity's Last Exam

41.4%

LMArena Elo

1496

MMLU-Pro

—

SWE-bench Verified

—

AIME (latest)

—

ARC-AGI-2

—

Strengths & weaknesses

Strengths

Takes text, images, video and documents in one system — the first Qwen flagship to do all four
Available same-day inside Alibaba's own coding and agent platforms

Weaknesses

A preview described as continuously evolving — behavior is a moving target, not a fixed release
No model card or active-parameter count published, so its real speed and cost profile are unknown
Reachable only through subscription plans rather than a per-token endpoint

Announced 2026-07-19 at WAIC Shanghai as a preview, then given a general-availability launch with standard API pricing on 2026-08-03 (Alibaba Cloud Model Studio blog: "Qwen3.8-Max: A New Bar for Coding and Cowork"). Alibaba's own release benchmark table (self-reported, published as an image and not independently re-derivable) claims GPQA Diamond 92.6, Terminal-Bench 2.1 86.6, HLE 43.6, and SWE-bench Pro 67.7 (not the SWE-bench Verified variant this site tracks, so left unset here) — relayed via MarkTechPost and OfficeChai coverage of the announcement. Artificial Analysis's own independent evaluation (artificialanalysis.ai/models/qwen3-8-max and its per-benchmark leaderboards) measured GPQA Diamond 92.7 (close to Alibaba's claim) but Terminal-Bench v2.1 81.3 and HLE 41.4 — both notably below Alibaba's self-reported figures — so the AA-measured numbers are recorded here as the more independently verifiable source, with the discrepancy noted. Pricing ($2.00/$6.00 per 1M input/output tokens) is Alibaba Cloud Model Studio's Singapore/international rate; the China Beijing, Frankfurt, Virginia, Tokyo and Hong Kong regions price lower at $1.65/$4.951. Open weights, promised "next week" as of the Aug 3 announcement, arrived on 2026-08-13 — but as `qwen3-8-2-4t-a95b`, a reduced text-only variant with a 262K native window rather than this multimodal 1M-context Max build, which stays closed. Context window (983,616) and max output (131,072) match Alibaba Cloud Model Studio's official model-info page (context 1,000,000, max input 991,808 with 1,000,000 minus max-output overhead, max output 131,072); some coverage rounds context down to "1M tokens."

For developers

API model strings

Alibaba Cloud Model Studioqwen3.8-max

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor