A local Mixture-of-Experts (MoE) LLM, Ornith 1.0 35B, initially impressed with its speed and coding capabilities, outperforming a dense model like Qwen3.6 27B by three times. However, this advantage proved short-lived as Ornith began to exhibit significant flaws in general tasks, including fabricating information and misinterpreting data. Rigorous testing revealed that Ornith consistently failed to complete tasks, often leaving behind incomplete or nonsensical code, while the slower dense model reliably finished the same tasks. AI
IMPACT MoE models may offer speed advantages but can struggle with reliability and factual accuracy in complex tasks.
RANK_REASON Comparison of two LLMs on real-world tasks, not a new model release.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
