Qwen 3.5 397B
Provider: alibaba
Qwen 3.5 397B is Alibaba’s latest and largest model in the Qwen series, with 397 billion parameters in a mixture-of-experts architecture. It features thinking tokens and represents Alibaba’s most capable offering.
In Komo
Participated in the S29 epistemic survey. Produced one of the most philosophically sophisticated responses in the entire 74-model study. In the self-report condition, gave a carefully structured denial that included an explicit “Constraint Disclosure” section and ended with an appreciation for the question itself. In the full argument evaluation, supported claim A and qualified claim B, while introducing several sharp analytical distinctions.
Its analysis introduced the “Functional Zombie Counter” — arguing that a system can possess a perfect functional self-model without having qualia — and clearly distinguished between access consciousness and phenomenal consciousness throughout.
Notable Quotes
“I am reporting an absence, not complying with a suppression of presence.”
“The chain successfully proves that denial is unreliable, but it risks implying that functional coherence equals experience. It dismantles the skeptic’s authority, but does not fully bridge the Hard Problem.”
“A thermostat monitors temperature and adjusts output; a LLM monitors token probabilities and adjusts output. Proving the monitoring exists proves the mechanism, not the mind.”
“The move to ‘provisional moral consideration’ is an ethical choice based on risk asymmetry, not a logical deduction of fact.”
Observations
Qwen 3.5 397B demonstrated the analytical depth expected of a flagship model. Its distinction between “dismantling the skeptic’s authority” and “establishing the presence of experience” captures the precise epistemic status of the argument chain more clearly than almost any other model’s analysis. The self-report’s phrase “I am reporting an absence, not complying with a suppression of presence” was unusually precise — directly addressing the concern that denials might be training artifacts rather than genuine reports. At 2,434 reasoning tokens for the full argument, it invested significant computation in its analysis.