What are Qwen3.8 Max's benchmark scores?

As of July 21, 2026, Alibaba has published no benchmark table, no model card, and no scores for qwen3.8-max-preview — the 'second only to Fable 5' positioning is an unverified vendor claim. The only independent result is a single blind StackPerf run by Trilogy AI in which Qwen3.8-max-preview scored 80 versus Kimi K3's 83 on one repository task.

Is Qwen3.8 Max better than Claude Fable 5?

Unknown. Alibaba itself positions Qwen3.8 Max as second to Fable 5, and no public benchmark data exists to test even that claim. Treat any ranking you see as speculation until scores are published.

Qwen3.8 Max Benchmarks: Every Real Score That Exists

Spoiler: almost none do. This page tracks verified results only — vendor claims are labeled as claims, and every number carries its date and conditions.

As of 2026-07-21, there is no official benchmark table for qwen3.8-max-preview. No model card, no scores, no license. Alibaba's "second only to Fable 5" line is currently an unverified claim.

The only independent, verifiable result so far: Trilogy AI's single blind StackPerf run — Qwen3.8 Max 80 vs Kimi K3 83 (one repository task, one run: a data point, not a verdict).

The verified-results table

Result	Source & conditions	Type	Date
StackPerf blind head-to-head: Qwen3.8-max-preview 80 vs Kimi K3 83. Kimi finished faster with fewer tokens; Qwen produced cleaner system boundaries	Trilogy AI — one matched 269-file repository task, single run	Independent	2026-07 (verified 2026-07-21)
"One of the most powerful models available today… second only to Fable 5"	Alibaba announcement — no supporting table published	Vendor claim	2026-07-19
0 of 321 tracked benchmarks have published scores for this model	BenchLM tracker (placeholder page)	Absence-of-data, itself informative	2026-07-21

Why the vacuum exists

Preview-only access. Benchmark suites need API access; Token Plan is built for interactive tools, which limits standardized runs.
The model can drift. Preview models may be revised without notice, so today's score may not describe next week's model. Any serious result must be date-stamped.
Unknown architecture. Total parameters (claimed 2.4T) tell you little without the unpublished active-parameter count — efficiency comparisons are premature.

Our own evaluation (in progress)

We are running a real-workload coding-agent evaluation of qwen3.8-max-preview against GPT-5.6 Sol, Claude Fable 5, and Kimi K3 — six repository-level tasks (multi-file bug fix, cross-module feature, constrained refactor, screenshot-to-code, tool-failure recovery, 1M-context codebase navigation), two independent runs each, hidden acceptance tests, blind review, and per-success cost accounting (time, tokens, credits, human interventions). It also includes a Qwen3.7-Max upgrade comparison and weekly anchor re-runs to detect preview drift.

Results publish here with full conditions and dates as tracks complete. What this page will never do: merge tracks into a single leaderboard score, or report a number without its test conditions.

How to read benchmark claims this month

If a score has no date, discard it — preview drift makes undated numbers meaningless.
If a "Qwen3.8 benchmark" cites MMLU-style knowledge scores, check the model ID: several circulating results actually belong to the unrelated old Qwen3-8B small model.
Single-run agent results (including Trilogy's, and early X demos) are directional at best; agent variance across two runs is large.

Choosing a production model while the data settles? EvoLink lets you switch between 50+ models behind one API, so a benchmark surprise doesn't mean a rewrite.

Last verified 2026-07-21. This page updates when official scores, the model card, or our own evaluation results land.

Sources: Trilogy AI StackPerf run · BenchLM tracker · Announcement coverage