Qwen 3.8 Max (Preview)

Spec

Attribute	Value
Developer	Alibaba / Qwen AI Lab
Released	not yet (preview only as of 2026-07-19)
Announced	2026-07-19
Context window	unknown
Pricing	unknown
License	Open-weight (planned; no firm date)
Availability	Preview: Alibaba Cloud Token Plan, Qoder, QoderWork

Release Date

Preview announced July 19, 2026. Full open-weight release: no firm date.

Benchmarks

None published. Alibaba's self-claim: "second only to Anthropic Fable 5." No independent third-party benchmarks as of 2026-07-19 — in contrast to Qwen 3.7-Max (May 2026), which cited Terminal Bench, SWE-bench Pro, and long-horizon demos.

Architecture

• 2.4 trillion total parameters (MoE)
• Active parameter count at inference: not disclosed — a critical gap for assessing true inference cost
• Architecture details not yet published

Context

• Announced a few days after Kimi K3 (2.8T, July 16)
• Makes Qwen 3.8 the second-largest publicly previewed model behind Kimi K3
• The claim of "second only to Fable 5" rests entirely on Alibaba's internal evaluations

Sources

• Bloomberg: Alibaba's Qwen Unveils Preview
• MLQ News: 2.4T parameters