This week saw several significant developments in large language model deployment and evaluation. xAI released Grok CLI, an open-source agentic coding tool powered by Grok 4.6 that utilizes xAI's free API tier, potentially shifting the economic landscape for usage-based agentic sessions. Alibaba Cloud introduced Qwen3.8-2.4T-A95B, a massive sparse Mixture-of-Experts (MoE) model released under the Apache 2.0 license; it features 2.4 trillion total parameters but only utilizes 95 billion active parameters per forward pass to manage computational demands. In the realm of optimized inference, Z.AI launched GLM-5.2 Turbo, a speed-optimized variant of the GLM-5.2 base designed to improve throughput and latency for multi-step agentic coding tasks through distillation or quantization. Additionally, the deepseek-ai/deepseek-harness gained significant traction on GitHub, reaching over 20,000 stars, as a reproducible evaluation framework for DeepSeek model variants. Finally, the emergence of ASI-Bench on HuggingFace signals a shift in benchmarking methodology, aiming to differentiate frontier models at the capability ceiling where traditional metrics like MMLU and HumanEval have reached saturation and lack discriminatory power.