Overview This article compares the resource requirements and usage costs of three major domestic LLMs, helping developers choose the right solution for their scenarios.
Model Vendor Architecture Minimum Deployable VRAM API Available GLM-5 Zhipu AI Dense (multiple versions) 24GB (8B) ✅ Kimi K2.5 Moonshot AI MoE (undisclosed) 24GB (lightweight) ✅ MiniMax M2.7 MiniMax MoE 230B Not yet open-sourced ✅ GLM-5 (Zhipu AI) Versions & Hardware Requirements GLM-5 offers 4 parameter versions, making it the widest-coverage domestic LLM currently available.
What is Harness Engineering? Definition: Harness Engineering is the discipline of designing constraints, feedback loops, tool systems, and verification mechanisms around AI agents.
This definition can be understood through an analogy:
Harnessing a Thousand-Mile Horse: A thousand-mile horse (AI Agent) is capable of running fast, but without a rider, it might run randomly, injure passersby, or even rush off a cliff. Harness Engineering equips this horse with reins (constraints), brakes (safety controls), whip (incentive mechanisms), and a rider (monitoring), ensuring it travels safely on the correct path.
Scenario: Information Is Correct, But Execution Goes Wrong A company deployed a RAG-based technical documentation Q&A system. The system worked well—when users asked “How to configure Redis cluster?” it could accurately retrieve relevant information from technical documents and provide detailed configuration steps.
Problem: When a user asked “Delete temporary files in the test directory,” the system correctly retrieved the right technical documentation, but during execution it mistakenly deleted the entire project’s core code.
What is Context Engineering? In June 2025, Andrej Karpathy provided a definition of Context Engineering on the OpenAI engineering blog: “the delicate art and science of filling the context window with just the right information for the model to take the next step.”
This definition captures the core distinction from Prompt Engineering:
Prompt Engineering: Optimizes “what you say” – focuses on how input instructions are expressed Context Engineering: Optimizes “what the model knows” – focuses on what information the model can access Using a chef as an analogy: Prompt Engineering adjusts the menu instructions given to the chef, while Context Engineering manages the complete ingredient warehouse the chef can draw from.
The Problem Consider a prompt on “the latest best practices for Python MySQL connections,” with a carefully designed persona (“You are a Python database expert with 10 years of experience”), clear instructions (“Only provide 2024 best practices, no deprecated methods”), and specific format requirements (“List main approaches, pros/cons, code examples, security considerations”).
The prompt itself is well-designed, yet GPT may still confidently return code using 2018-deprecated methods with security vulnerabilities. The reason: the model doesn’t know 2024’s best practices, because its training data cutoff doesn’t include this new information. The problem isn’t the prompt—it’s the model’s knowledge.
What is Prompt Engineering? The core definition of Prompt Engineering is: Designing natural language inputs to guide Large Language Model outputs toward specific results.
This concept seems simple, but it hides a profound assumption: The same model, different prompts → completely different outputs.
Imagine you have an incredibly smart assistant with zero background knowledge. This assistant can perfectly understand and execute any instruction, but it lacks prior knowledge and has no memory. Prompt engineering is the art of learning how to converse with such an assistant.