Practical AI learningTechnical / models
Free public lessons

Technical / models

Models. Open weights, APIs, cutoffs.

Compare three published checkpoints, verify current API choices with provider documentation, and understand what a training cutoff can and cannot tell you.

SOURCE STATUS SEE EACH TABLE // TABLES 3 // SORTABLE CLICK ANY COLUMN HEADER // VERIFY BEFORE CITING
ON THIS PAGE
  1. Three published checkpoints to compare
  2. Choose an API using current provider evidence
  3. Read a cutoff without overclaiming accuracy
TABLE // OPEN WEIGHTS CHANGES QUARTERLY

Three published checkpoints to compare

Examples for learning to read a model card. These are not deployment recommendations or an exhaustive model catalog.

Linked source facts checked 2026-10-02; no deployment approval or performance validation implied.

Exact checkpoint Published size Published context License label What to check
meta-llama/Llama-3.3-70B-Instruct 70B 128K tokens Llama 3.3 Community License Read the exact license and acceptable-use terms; test your task and serving configuration.
microsoft/phi-4 14B (model-card description) 16K tokens MIT Text model; check intended use, language limitations and hardware requirements.
Qwen/Qwen3-8B 8.2B 32,768 native; 131,072 with YaRN configuration Apache 2.0 Extended context requires configuration. Test quality and resource use for your actual input length.
TABLE // API SELECTION CHANGES MONTHLY

Choose an API using current provider evidence

Before a live exercise, choose an available model in your own account and record its exact ID, endpoint, date and costs. The offline B3 lab requires no account or spending.

Linked source facts checked 2026-10-02; no deployment approval or performance validation implied.

Provider Read before running Record for reproducibility
OpenAI Official model catalog, Responses API documentation and pricing Exact model ID; input/output limits; storage setting; enabled tools; timeout/retry policy.
Anthropic Official Claude model overview, Messages API documentation and pricing Exact model ID; max_tokens; supported features; timeout/retry policy; data-handling terms.
TABLE // DATA FRESHNESS CHANGES QUARTERLY

Read a cutoff without overclaiming accuracy

A training cutoff describes a dataset boundary. It does not guarantee that a model knows an earlier fact or can verify a current claim.

Linked source facts checked 2026-10-02; no deployment approval or performance validation implied.

Checkpoint Published statement Practical implication
Llama 3.3 70B Instruct Pretraining cutoff: December 2023; released December 6, 2024 Release date and pretraining cutoff are different. Verify current facts with authoritative sources.
Phi-4 Public-data cutoff: June 2024 and earlier; released December 12, 2024 The card describes several training-data types. A single cutoff does not describe every fact learned.
Any model with retrieval or browsing Retrieved material may be newer than training data Check the retrieved source, date and supported claim; retrieval does not itself establish truth.

A source-check date applies only to the facts in that table. Tables without a date retain older orientation data and have no recorded source verification date. Verify current prices, versions and capabilities with primary sources before spending or citing them. Licenses and brand names do not establish authorization.