Qwen3.6 35B fails Stargate counting test

Qwen3.6 demonstrates that "frontier" reasoning models still hit a wall with basic character-level awareness and linguistic groundedness.

• –The model hallucinated the spelling of "O'Neill" as "O'Neil" to justify a single-L count, showcasing a classic conflict between internal knowledge and token-based logic.
• –Sparse MoE architectures like the 35B-A3B variant optimize for inference speed but may sacrifice the granular precision needed for non-textual reasoning tasks.
• –The gap between its 92.7% AIME score and its failure on a "vibe check" counting task highlights the limitations of current synthetic benchmarks in predicting real-world reliability.
• –While the model is highly corrigible through multi-turn dialogue, its zero-shot reasoning remains susceptible to the same pitfalls as previous LLM generations.