Discofork
Repository brief

karpathy/nanochat

Read the upstream summary on the left, browse the cached forks below it, and load each fork comparison into the right-hand panel.

Cached analysis
cached 2026-03-30T12:35:52.728Z
3mo ago

karpathy/nanochat

karpathy/nanochat is a Python-based, single-GPU experimental harness for training and chatting with LLMs, positioned as a minimal, hackable full-stack ChatGPT clone. It covers tokenization, pretraining, finetuning, evaluation, inference, and a web chat UI, and is currently optimized around a GPT-2 speedrun benchmark with active development focused on pretraining efficiency.

GitHub
Loading tags...
Stars50,662
Forks6,644
Default branchmaster
Last pushed2026-03-27T13:44:17Z
Recommended shortcuts

Jump straight into Discofork's strongest cached fork picks, or open a compare view in one click.

Forks

Choose a fork to inspect

10 of 10 fork briefs
Selected

Prefer this fork if your priority is cloud deployment and checkpoint logistics; prefer upstream if you want the current benchmark-focused nanochat line and the latest training improvements.

Prefer upstream unless you specifically need this older, unchanged snapshot. This fork adds no visible capabilities, while lagging 345 commits behind upstream means it likely misses newer fixes and training improvements.

Choose upstream unless you specifically want this fork’s exact pinned state; it adds no visible capability and is already materially behind active upstream work.

Choose this fork if your priority is deployment and demo workflows across hosted platforms; choose upstream if you want the latest training-efficiency work, benchmark tuning, and a smaller moving target.

Choose the fork only if you specifically want a frozen older baseline; otherwise upstream is the better choice because this fork is materially stale and provides no evident added capability.

Choose the upstream repo unless you specifically need this older checkpoint; this fork adds no visible capabilities and is materially behind on active development.

Choose upstream instead unless you specifically want this fork as a frozen, unmodified snapshot. It adds no visible features and trails upstream by 363 commits.

Choose this fork if you want an EBT-focused research branch with substantial new training and model code. Choose upstream instead if you want the current minimal ChatGPT-style harness, the latest benchmark optimizations, and the most active reference implementation.

Prefer this fork if your goal is MoE experimentation and richer training/export workflows. Prefer upstream if you want the current, simpler nanochat benchmark-oriented stack with less maintenance risk and better compatibility with ongoing upstream improvements.

Prefer upstream unless you specifically need this exact historical snapshot; the fork adds no new capability and is substantially behind current nanochat.