Alibaba (Qwen)

Qwen3-235B-A22B

Released Apr 28, 2025 · knowledge cutoff 2024-10

Status: Superseded
Location: China
Type: LLM
Modality: Text
Context window: 131K
Max output: 33K
Speed: —
Price ($/MTok in / out): $0.20 / $0.60
Cost per task: —
Open weights: Yes

Benchmarks6 of 10 reported

MMLU-Proretired

82.8%

GPQA Diamond

71.1%

SWE-bench Pro

21.41%

AIME (latest)retired

85.7%

Humanity's Last Exam

11.8%

LMArena Elo

1373

SWE-bench Verified

—

Terminal-Bench 2.1

—

GDPval-AA v2

—

ARC-AGI-2

—

Strengths & weaknesses

Strengths

Switches between instant answers and extended thinking inside one model
Very broad language coverage, including low-resource languages most models drop
Permissive licence with no regional restrictions

Weaknesses

Loses coherence on long agentic runs where closed models hold together
Thinking mode is verbose and hard to bound

Scores for the July 2025 updated (2507) checkpoints where applicable. SWE-bench Pro 21.41% is Scale's public-set leaderboard figure for "qwen3-235b-a22b" — third-party, not Alibaba-reported; Scale does not state which Qwen3-235B-A22B checkpoint (original or 2507) it ran.

For developers

API model strings

Not researched

Licence

Apache License 2.0 · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor

News