This content evaluates the performance of the Qwen 3.6 35B A3B Q4_K_M quantized MoE model on CPU, using benchmarks like HumanEval, HellaSwag, and BFCL. It achieved 22 tokens/sec, showing strong performance in commonsense reasoning (74%) and solid results for an active 3B MoE model.
This article evaluates Claude Opus 4.6, Sonnet 4.6, and Haiku 4.5 across 10 real-world tasks using AgentHunter Eval v0.3.1. Opus 4.6 was the only model to achieve a perfect score, while Haiku 4.5 proved significantly faster and more cost-effective despite failing one task alongside Sonnet 4.6.
The 'review-prompts' GitHub repository offers an open-source collection of prompts to improve AI model evaluation, aiming for comprehensive understanding of performance, biases, and limitations. This resource fosters a more rigorous and transparent approach to AI development for the builder community.