Alibaba (Qwen)

Qwen3.5

Released Feb 16, 2026 · knowledge cutoff unpublished

Status: Unknown
Location: China
Type: LLM
Modality: Text
Context window: 262K
Max output: 66K
Speed: 82 tok/s · 2.41s to first answer token
Price ($/MTok in / out): $0.54 / $3.40
Cost per task: $0.33
Open weights: Yes

Benchmarks3 of 10 reported

MMLU-Proretired

87.8%

GPQA Diamond

88.4%

SWE-bench Verified

76.4%

SWE-bench Pro

—

Terminal-Bench 2.1

—

AIME (latest)retired

—

Humanity's Last Exam

—

LMArena Elo

—

GDPval-AA v2

—

ARC-AGI-2

—

Strengths & weaknesses

Strengths

Very few active parameters per token, so it serves cheaply for its capability
Open weights that fine-tune well and run on commodity inference stacks
Strong multilingual coverage carried over from the Qwen line

Weaknesses

Text-only, so it can't take screenshots, diagrams or PDFs
No first-party API — behavior and speed depend entirely on which host you pick

Figures are for the Qwen3.5-397B-A17B open release. MMLU-Pro 87.8 / GPQA 88.4 / SWE-bench 76.4 third-party (LLM-Stats/DeepInfra). Pricing is the DeepInfra hosted FP8 rate — no first-party API for the open weights. Speed per Artificial Analysis' Qwen3.5 397B A17B (Reasoning) page; AA labels the variant by reasoning mode, not an effort tier.

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