Weights & Biases

What Weights & Biases offers learners

Weights & Biases is an AI developer platform that teams use to train, track, evaluate and monitor AI models and agents, and its AI academy is a set of free courses for the engineers who do that work. The courses page opens with "Learn to use foundation models and agents in your AI applications. Implement MLOps and LLMOps solutions." and describes the material as free courses "trusted by thousands of practitioners."

This entry covers the learning side. The reader it serves is a developer, data scientist or machine learning engineer who wants a short, practical course on one production problem: building an agent, evaluating an LLM app, getting a RAG system out of the prototype stage, fine-tuning a model, or putting continuous delivery around a model. Every course on the page carries a Free label and an "ENROLL for free" button, so the catalog costs nothing to take, and the same page links to sign-up and log-in for the platform itself.

The course catalog

The catalog is small and specialized rather than broad. The courses page filters it into GenAI, MLOps, W&B Platform and International, and gives each course a short description and a running time.

The GenAI courses carry most of the weight:

• AI Engineering: Agents (2 hours) teaches learners to work with reasoning models and build production-grade agents that "autonomously tackle complex workflows, from deep analysis to support agents", and to architect and evaluate systems that combine LLMs, tools, memory and planning. The page promotes it as its newest GenAI course, "created in collaboration with OpenAI team."
• LLM apps: Evaluation (2 hours) covers building, optimizing and scaling AI evaluators with minimal human input, and building evaluation pipelines that combine programmatic checks with LLM-based judges.
• RAG++: From POC to production (2 hours) is about production-ready retrieval-augmented generation, with the stated aims of better performance, lower cost and more accurate, relevant results.
• Developer's guide to LLM prompting (2 hours) runs from system prompts and structural techniques to model-specific strategies, worked through a text-understanding use case.
• LLM engineering: Structured outputs (2 hours) is taught with Jason Liu, whom the page names as the author of the Instructor library, and covers structured JSON output, function calling and complex validations with Pydantic.
• Building LLM-powered apps (3 hours) walks through designing, experimenting with and evaluating LLM-based apps using LLM APIs, Langchain and the company's own LLM tooling.
• Training and fine-tuning LLMs (4 hours) covers the history and architecture of LLMs, training techniques and fine-tuning methods such as LoRA and Prefix Tuning, with practical recipes from Jonathan Frankle (MosaicML) and other practitioners. The page says it suits machine learning engineers, data scientists, researchers and NLP enthusiasts.

The MLOps courses deal with the operational side of models:

• Model CI/CD (2.5 hours) covers automating key workflows, governance and the end-to-end model lifecycle.
• Effective MLOps: Model development (4 hours) is about building a disciplined, collaborative process and an end-to-end pipeline that keeps delivering production-ready models.
• CI/CD for machine learning (GitOps) (5 hours) is the longest course; the page says learners "use GitHub Actions and integrate W&B experiment tracking" to automate pipelines and deploy models.
• Data validation in production ML pipelines (2 hours) covers data validation, detecting data drift and managing data quality with automated toolkits.

Under the International filter sits a Japanese edition of the RAG++ course, titled RAG++ : PoCから本番運用へ, listed as two hours of content (2時間のコンテンツ) and marked free (無料).

How taking a course works

Access is simple. A learner picks a course, presses the enroll button and starts; the courses page shows no fee, tier or paywall for any of them. The running times make the format clear enough: two hours for most courses and three to five for the longer ones, so a working engineer can finish one in an afternoon or spread it over a week of evenings. The descriptions stress hands-on work: the GitOps course calls itself a practical, hands-on learning experience, the prompting course works through code on a real use case, and the fine-tuning course promises hands-on recipes.

Several courses teach with the company's own tools. The LLM app course builds with Weights & Biases LLM tooling and the GitOps course wires in W&B experiment tracking, so a learner finishes with working knowledge of the platform as well as the topic. The home page's quick-start examples are Python (import wandb, weave.init), which is the language a learner should expect to read. The courses page does not mention a certificate of completion, so the value here is the skill rather than a credential.

The platform behind the courses

The courses make more sense next to the product they come from. The home page groups the platform into a few parts:

• Models: Experiments to track and visualize runs, Sweeps for hyperparameter optimization, Tables to explore data and Reports to document and share results.
• Weave: Traces to explore and debug AI applications, Evaluations, a Playground for prompts and models, and Monitors for improving apps in production.
• Serverless training: Serverless RL to fine-tune LLMs without managing GPUs, Serverless SFT to teach LLMs new tasks, the open-source RL framework ART and Ruler, an automated reward function for RL.
• Serverless inference: hosted models including GPT OSS 20B and 120B, Qwen3 235B A22B, 235B Thinking and Coder 480B, Llama 4 Scout, Llama 3.3 70B and Llama 3.1 8B, Kimi K2.5, Phi 4 Mini 3.8B, DeepSeek V3.1, V3-0324 and R1-0528, and Z.AI GLM 5.0.
• Core: a Registry to publish and share models and datasets, Skills for coding agents, CoreWeave Sandboxes as isolated environments to run agents, Automations, and ARIA, described as an AI Research and Iteration Agent.

Under "INTEGRATE QUICKLY" the home page lists LangChain, LlamaIndex, PyTorch, Lightning, TensorFlow, Keras, Scikit-learn and XGBoost, each with a short code sample. The use cases it names run from AI governance, computer vision and contact centers to physical AI, RAG, training LLMs and quant trading, and it links case studies from Canva, Microsoft and Toyota. There is also an iOS app, which the home page calls "The first iOS app to monitor AI experiments and track training runs anytime, anywhere."

One customer quote on the home page, from Mike Maloney, Co-founder and CDO at Neuralift, sums up the pitch for Weave: "The fact that I could just add a library to our code" and get information about the GenAI part of a product inside a tool the team already used.

Security, deployment and support

For teams that move from the courses to the platform, the home page states that the platform "is certified under ISO/IEC 27001:2022, ISO/IEC 27017:2015, and ISO/IEC 27018:2019, and is compliant with SOC 2 and HIPAA standards", that it helps customers comply with NIST 800-53, and that the company is in alignment with GDPR requirements for processing personal information. Deployment comes as multi-tenant SaaS, a dedicated instance or a customer-managed install. The site offers sign-up and a demo request, and gives support@wandb.com as the customer support address. The site itself switches between English, Japanese, Korean and German.

Who it suits and who it does not

The courses suit engineers and data scientists who already build with machine learning and want to get better at one specific part of shipping it: agents, evaluation, RAG, structured outputs, fine-tuning or model CI/CD. They fit teams that use Weights & Biases or are evaluating it, since the exercises use its tools, and they fit anyone who wants focused material for free without a subscription. Japanese-speaking engineers get one course in their own language.

They are a weaker fit for someone new to programming or to machine learning, because the topics assume working knowledge of models, pipelines and Python. A learner who wants a broad, structured program with a credential at the end, or live teaching and mentor feedback, will find the catalog narrower than a full course provider; this is a focused set of short courses from a tools company, and on those terms it is generous.