Atlas

Models

← All models

Qwen3.5-27B

Alibaba · 2026-02-24 · 27.8B parameters

Qwen3.5-27B is Alibaba's open-weight, dense, post-trained multimodal model in the Qwen3.5 medium series, released February 24, 2026. Its complete checkpoint contains 27,781,427,952 parameters; it accepts text, images, and video, produces text, uses a hybrid Gated DeltaNet/full-attention architecture, supports optional thinking, and is Apache 2.0 licensed. It supports 262,144 tokens natively and documents YaRN extension to 1,010,000 tokens.

Benchmark scores

BenchmarkScore
AA-LCR67.3
AA-LCR v1.177.7
AA-Omniscience Index-44.0
AgentVQA46.7
AgentVQA Egocentric Videos57.0
AgentVQA Games49.5
AgentVQA Robotics43.0
AgentVQA Spatial Planning51.5
AgentVQA Web Agents32.5
AIME 202691.7
ALE-Bench882
Apex (MathArena)2.1
Apex Shortlist52.1
Artificial Analysis Omniscience Accuracy20.7
Artificial Analysis Omniscience Hallucination Rate75.1
Artificial Analysis Openness Index38.9
Artificial Analysis Progress Index v1267
ArXivMath 01/202653.3
ArXivMath 02/202631.3
ArXivMath 12/202541.2
BuseyBench SVG2.8
Capability147.1
CritPt (avg@5)0.9
CulturalRiddles28.0
CulturalRiddles Exact Match13.8
DTBench82.4
Eval Connections60.0
Eval Connections — Guess Accuracy57.7
Extended NYT Connections 940 (Quadratic v1)47.9
Extended NYT Connections 940-Puzzle Set60.8
Loading Atlas data…