Gemma 4 hallucinates Qwen 3.5 claims

This looks less like secret 2026 training and more like classic LLM overconfidence: a model can surface a real-sounding name and still be wrong about the facts around it.

• –Qwen3.5 is an actual Qwen release, so the answer may be anchored to a real model name rather than pure fiction
• –Confidence is not evidence of freshness; a model can be stale, contaminated, or simply hallucinating with high fluency
• –Quantization changes quality and reliability, but it does not add new knowledge
• –For local model testing, factual grounding and calibration matter more than a single impressive response
• –The post is a good reminder that cutoff dates are not a hard guarantee against later-sounding outputs