MODEL ID: qwen3.8-max-preview
STATUS: PREVIEW LIVE
WAIC SHANGHAI
UTC --:--:--
UNVERIFIED LAUNCH
Alibaba CloudQwen TeamPreviewed 19 Jul 2026

Qwen 3.8Max · Preview

Total parameters — Alibaba's own figure
0
sparse mixture-of-experts · active-parameter count undisclosed

Two days after Moonshot's Kimi K3 rattled chip stocks, Alibaba walked onto the WAIC stage with a bigger number of its own — "second only to Fable 5." This dossier separates what the lab confirmed from what it only claimed.

01 · The verdict

Confirmed vs. claimed

The preview is real and purchasable. The headline numbers are not — yet. Every performance figure at this launch carries a caveat.

✓ Confirmed

5 items

Preview is live and purchasable — sold through Alibaba's Token Plan, Qoder and QoderWork at 10% of standard pricing.Qwen announcement · MarkTechPost

Sparse Mixture-of-Experts, multimodal — the team's first model above 1T parameters handling text, images, video and documents.Shuai Bai, Qwen developer

OpenAI & Anthropic protocol compatible — existing coding agents can point at it without rebuilding their harnesses.Alibaba team statement

Qwen3.7-Max baseline is verified — GPQA-Diamond 92.4, SWE-bench Verified 80.4%, 1M-token context, broadly corroborated by Artificial Analysis.Alibaba · Artificial Analysis

Open-weight release announced — the Qwen account says Qwen3.8 is going open-weight "soon".@Alibaba_Qwen

✗ Claimed only

5 items

2.4 trillion parameters — Alibaba's own figure. No model card, no spec sheet, no independent check.WAIC announcement

"Second only to Fable 5" — rests entirely on internal evaluation. No benchmark table was published.Qwen official account

Beats 3.7-Max on coding, full-stack, data analysis — the team states it "should" outperform. No scores accompany the claim.Shuai Bai

Active-parameter count — the single number that determines real inference cost for a sparse MoE. Undisclosed.not published

Open weights delivered — the previous two Max-tier flagships also launched closed and never open-sourced.AI Tools Review lineage analysis

02 · The numbers

Benchmarks: what's on paper

The only verified scores belong to the predecessor. For 3.8 itself, the bar is still empty.

SOURCE: Qwen3.7-Max — Alibaba-published, broadly corroborated by Artificial Analysis (AA Intelligence Index 56.6, 5th overall at release)
GPQA-Diamond92.4
graduate-level science reasoning · Qwen3.7-Max
SWE-bench Verified80.4%
real-world software engineering · Qwen3.7-Max
Terminal-Bench 2.0 · Terminus69.7
agentic terminal tasks · Qwen3.7-Max
Qwen3.8-Max-PreviewAWAITING INDEPENDENT SCORES
no benchmark table, model card or third-party evaluation published at preview time
03 · The race

Three trillion-class releases in one month

GLM 5.2, Kimi K3, Qwen 3.8 — the Chinese open-weight surge, and the in-house rivalry behind it: Alibaba owns roughly 36% of Moonshot AI.

Sep 2025
Qwen3-Max-Preview

Qwen crosses the trillion-parameter threshold for the first time.

Apr 2026
Qwen 3.6 Max Preview STAYED CLOSED

Hosted only through Qwen Studio and Alibaba Cloud — a departure from shipping openly on day one.

May 2026
Qwen3.7-Max

Incremental step: 1M-token context, GPQA-Diamond 92.4, SWE-bench Verified 80.4%.

Jun 2026
GLM 5.2 · Z.AI

Draws praise from Vercel and Box engineers for landing within striking distance of Claude Opus 4.8 on FrontierSWE.

17 Jul 2026
Kimi K3 · Moonshot AI OPEN WEIGHTS

2.8T-parameter open-weight launch briefly tips the Philadelphia Semiconductor Index toward a bear market.

19 Jul 2026 · WAIC Shanghai
Qwen3.8-Max-Preview

2.4T parameters, multimodal, "second only to Fable 5". Live and purchasable — weights promised, not delivered.

04 · Access

Token Plan pricing

Bundled credits, not per-token pricing. The preview runs at 10% of standard rates.

$6 /week
Entry tier
  • 2,500 credits per 7 days
  • Token Plan subscription
  • Preview discount applied
Caveat: only the two ends of the range are publicly documented. Credit-to-token conversion rates have not been published, so per-inference cost cannot be independently calculated yet. SOURCE: AI TOOLS REVIEW · ALIBABA TOKEN PLAN DOCS
05 · Context

Why the number alone says little

~36%
In-house rivalry

Alibaba owns roughly 36% of Moonshot AI — Qwen 3.8 vs. Kimi K3 is two portfolio companies trading blows in the same week.

~1.2 TB
Weight footprint

At 4-bit precision the full 2.4T model would need ~1.2 TB of storage — beyond any single multi-H200 node. What actually ships open is unknown.

22B / 235B
Sparsity precedent

Qwen3-235B-A22B activates only 22B of its 235B parameters per token. Without the equivalent ratio for 3.8, "2.4T" says little about serving cost.