Reinforcement Learning Interpretability Qwen

Qwen-Scope

Decoding Intelligence, Unleashing Potential

Open-source sparse-autoencoder interpretability toolkit for analyzing and steering Qwen3.5-27B model behavior, including repetition-related features and feature-overlap comparisons.

Pricing

Artificial Analysis · Jun 25, 2026, 1:04 PM

Input / 1M

$0.10

Output / 1M

$0.80

Blended / 1M

$0.275

Output TPS

263

TTFT (s)

1.08

Model Intelligence

Benchmarkable

Intelligence Index

GPQA

0.74

HLE

0.07

SciCode

0.26

IFBench

0.38

LCR

0.44

TerminalBench Hard

0.08

TAU2

0.85

Recent stories

1 linked story

releasePRIMARY2026-04-30

Qwen-Scope releases SAE toolkit for Qwen3.5-27B steering

Alibaba’s Qwen team released Qwen-Scope, an open sparse-autoencoder suite for Qwen3.5-27B that can steer outputs, surface repetition features, and compare benchmark feature overlap. The toolkit turns interpretability artifacts into debugging, data-generation, and evaluation workflows.