Contrast the unit of change that LLMOps ships against the unit MLOps ships

**MLOps** is the operational discipline that grew up around classical machine learning: train a model, version it with the data and code that produced it, deploy it, monitor it, retrain when it drifts. The artifact in the deploy is **model weights plus a feature-pipeline version**. Quality is gated on **offline evaluation on a held-out set**. **LLMOps** is the equivalent discipline for systems built on large language models. The base model is almost always a third-party black box (OpenAI's GPT-5.5, Anthropic's Claude Opus 4.7, Google's Gemini 3.1 Pro) or a pinned open-weights snapshot (Llama 4 served on vLLM). You do not retrain the base model; you rarely even fine-tune it. The artifact in the deploy is a **prompt diff, a tool definition, a retrieval index rebuild, a routing rule, or a vendor swap**. The gate is also different. Classical ML evaluates accuracy on a labeled held-out set. LLM systems evaluate via **eval as CI**: replay a golden set of representative inputs against the system and score outputs with a mix of rule-based checks (format validity, refusal rate, latency) and **LLM-as-judge** scoring against a rubric. This difference flows through everything: the team composition (more product engineers, fewer data scientists), the tooling (Langfuse, LangSmith, Promptfoo instead of MLflow, Weights and Biases), and the failure modes (prompt drift and retrieval-quality issues instead of feature drift and label decay).

**MLOps ships a trained model.** The unit of change is model weights, often co-versioned with the feature-pipeline code and the training-dataset snapshot. A deploy means swapping in new weights and a new feature pipeline, often gated on offline accuracy or AUC against a held-out test set. Retraining cadence varies by system but typically runs on a schedule (daily, weekly, monthly) or triggered by drift detection. Tools include MLflow, Weights and Biases, Kubeflow, Vertex AI Workbench, SageMaker, DVC. **LLMOps ships everything except the model.** The base model is rented from a hosted provider or pinned to an open-weights snapshot; retraining is rare and is almost never the path to iteration. The typical PR diff contains one of: - A **prompt edit** in the prompt registry (Langfuse Prompts, Anthropic's prompt management, a homegrown table) - A **tool definition change** (new tool added, schema modified, tool removed from the agent's toolkit) - A **retrieval index rebuild** (new documents indexed, chunking strategy changed, embedding model swapped) - A **routing rule** in the LLM gateway (LiteLLM, Portkey) that sends a different intent class to a different model - A **vendor or model snapshot swap** as part of a deprecation migration or quality upgrade The gate is **eval as CI**: replay a golden set of representative inputs through the full system, score outputs with a combination of rule-based checks (format validity, refusal rate, latency, cost per call) and LLM-as-judge scoring against a rubric. The CI artifact looks more like a test report than a model card; the deploy looks more like a config change than a model swap. **The disciplines overlap but are not interchangeable.** Both require versioning, observability, rollback paths, drift detection, and incident response.