AI Potluck
Model components / Fine-tuning code

LitGPT

Lightning AI

20+ from-scratch implementations of high-performance open LLMs (Llama, Mistral, Gemma, Phi, Qwen) with recipes for pretrain, finetune (LoRA, QLoRA, full), and deploy at scale. Picked by teams that want a clean, hackable, single-file-per-model codebase optimized for Lightning Fabric and FSDP, rather than the abstraction layers of transformers/TRL. 13.4K GitHub stars, 132 contributors; pace has slowed (14 commits in 90 days) but the library remains a popular teaching/research baseline.

LitGPT by Lightning AI, v0.5.12 released Dec 18 2025 (~6mo old at scoring; acceptable freshness). From-scratch LLM implementations with pretrain/finetune/deploy recipes for 20+ models. Confirmed live on GitHub June 2026.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI)
source
public
maintainer
Lightning-AI
core-gated
ungated

Apache-2.0, full source public; the library is unrestricted (Lightning's commercial offerings are separate products).

Adoption

3 medium confidence
3.0

~14.2k PyPI downloads last month (primary); 13.4k GitHub stars (corroborating only). Used in research (TinyLlama, NeurIPS 2023 LLM Efficiency Challenge starter kit). Modest but real recurring usage places it at the low end of medium; level 3 on monthly download volume, not stars.

Capability

3 medium confidence
3.0

Solid PEFT + pretraining recipes and clean from-scratch implementations, but narrower than verl/ms-swift/NeMo-RL (no RLHF/preference methods) and oriented to readable recipes over frontier-scale demonstration. Mid-tier within the category.

Unchanged since 2026-07-30 (last edited, not re-checked)