Open-project atlas¶
No single adjective captures an AI release. Use this atlas to choose a project by the artifact you need, then verify the current license and version at the source.
Checked: July 2026. “Published” means an official project makes the artifact available or provides a reconstruction path; it does not promise affordable reproduction, unrestricted licensing, or absence of data risk.
End-to-end research projects¶
| Project | Best reason to study it | Published artifact trail | Important boundary |
|---|---|---|---|
| OLMo 3 | staged base, mid-training, long-context, instruction, thinking, and RL flows | OLMo-core training scripts, checkpoints, Dolma 3/Dolci data curriculum, evaluations and model-flow artifacts linked by Ai2 | inspect the license and availability of each individual artifact; stages have different purposes |
| OLMo 2 | reproducible dense pretraining with configs, data links, checkpoints, and logs | training/model/eval code, official configurations, staged checkpoints, data files, linked run logs | older OLMo repository and newer OLMo-core represent different generations of the stack |
| OLMoE | open sparse-MoE study | weights, training code, data, evaluations, logs, and checkpoints linked by the project | small research MoE is a mechanism case study, not a proxy for every frontier MoE |
| Pythia | learning dynamics across sizes and steps | code, data reconstruction, controlled data order, and 154 checkpoints per main model described by the project | trained on versions of The Pile; study that corpus's provenance and availability separately |
| BLOOM | multilingual collaborative open science at large scale | model card, ROOTS data cards, training-code trails, optimizer/intermediate artifacts | weights use the BigScience RAIL license, which is not the same as a permissive software license |
| LLM360 Amber | release process and intermediate training artifacts | project links to model, data, code, configurations, metrics, and checkpoints | verify current mirrors and artifact-specific licenses before planning a reproduction |
Educational and implementation projects¶
| Project | Use it to learn | Not intended to prove |
|---|---|---|
| nanoGPT | a compact GPT-2-style model and training loop | an end-to-end modern frontier training stack |
| build-nanogpt | a commit-by-commit construction trace | data governance or large-scale post-training |
| LitGPT | clean implementations across model families, pretraining, fine-tuning, and inference | exact provenance of checkpoints trained by third parties |
| minbpe | byte-level BPE mechanics | production multilingual tokenizer engineering |
| llm.c | GPT-2 training and inference close to C/CUDA kernels | a general-purpose framework for all current architectures |
Build your component ledger¶
For any model family, fill this table with links, not checkmarks from memory:
| Component | Evidence | License/terms | Version | Status |
|---|---|---|---|---|
| Weights | available / partial / absent / unknown | |||
| Model code | ||||
| Training code | ||||
| Exact pretraining data | ||||
| Data processing | ||||
| Data mixture/order | ||||
| Tokenizer training | ||||
| Optimizer state | ||||
| Intermediate checkpoints | ||||
| Logs | ||||
| Post-training data/code | ||||
| Evaluation prompts/raw outputs |
Four questions hidden by “open source model”¶
- Can I run it? Weights, inference code, runtime support, hardware, license.
- Can I inspect it? Architecture, implementation, tokenizer, documentation.
- Can I reproduce it? Training data or sufficient data information, processing, training code, configuration, compute, state, and evals.
- Can I legally modify and redistribute it? Every relevant license and dataset term.
The OSI Open Source AI Definition 1.0 provides one formal standard for the freedoms and preferred form needed to modify an AI system. Project marketing terms do not override licenses.