{"id":49181297,"url":"https://github.com/rupeshs/awesome-microgpts","last_synced_at":"2026-04-23T01:31:02.114Z","repository":{"id":343910833,"uuid":"1177379509","full_name":"rupeshs/awesome-microgpts","owner":"rupeshs","description":null,"archived":false,"fork":false,"pushed_at":"2026-04-05T08:03:33.000Z","size":17,"stargazers_count":26,"open_issues_count":0,"forks_count":4,"subscribers_count":2,"default_branch":"main","last_synced_at":"2026-04-05T10:07:48.727Z","etag":null,"topics":[],"latest_commit_sha":null,"homepage":null,"language":null,"has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":null,"status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/rupeshs.png","metadata":{"files":{"readme":"Readme.md","changelog":null,"contributing":null,"funding":null,"license":null,"code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-03-10T01:09:41.000Z","updated_at":"2026-04-05T08:03:37.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/rupeshs/awesome-microgpts","commit_stats":null,"previous_names":["rupeshs/awesome-microgpts"],"tags_count":0,"template":false,"template_full_name":null,"purl":"pkg:github/rupeshs/awesome-microgpts","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/rupeshs%2Fawesome-microgpts","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/rupeshs%2Fawesome-microgpts/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/rupeshs%2Fawesome-microgpts/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/rupeshs%2Fawesome-microgpts/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/rupeshs","download_url":"https://codeload.github.com/rupeshs/awesome-microgpts/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/rupeshs%2Fawesome-microgpts/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":32162601,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-04-22T17:06:48.269Z","status":"ssl_error","status_checked_at":"2026-04-22T17:06:19.037Z","response_time":58,"last_error":"SSL_connect returned=1 errno=0 peeraddr=140.82.121.5:443 state=error: unexpected eof while reading","robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":false,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":[],"created_at":"2026-04-23T01:31:00.641Z","updated_at":"2026-04-23T01:31:02.102Z","avatar_url":"https://github.com/rupeshs.png","language":null,"funding_links":[],"categories":[],"sub_categories":[],"readme":"# awesome-microgpt [![Awesome](https://awesome.re/badge.svg)](https://awesome.re)\n\n\u003e A curated list of ports, variants, and extensions of Andrej Karpathy's [microgpt.py](https://gist.github.com/karpathy/8627fe009c40f57531cb18360106ce95) — a 200-line, dependency-free GPT in pure Python.\n\n---\n\n## Original\n\n| Name | Author | Language | Description |\n|------|--------|----------|-------------|\n| [microgpt.py](https://gist.github.com/karpathy/8627fe009c40f57531cb18360106ce95) | [@karpathy](https://github.com/karpathy) | Python | The original. 200 lines, zero dependencies, complete GPT: tokenizer, autograd, transformer, Adam, training loop, inference. MIT licensed. |\n\n---\n\n## Language Ports\n\n### Systems Languages\n\n| Name | Author | Language | Speedup vs CPython | Notes |\n|------|--------|----------|--------------------|-------|\n| [microgpt.cpp](https://github.com/Charbel199/microgpt.cpp) | [@Charbel199](https://github.com/Charbel199) | C++ | ~219x | Single-file, single-threaded. Tape-based autograd. SoA arena allocator. Fused multiply-add. Stack KV cache. Benchmarked ~109x vs PyPy. |\n| [rust-microgpt](https://github.com/mplekh/rust-microgpt) | [@mplekh](https://github.com/mplekh) | Rust | ~440x | Tape autograd with MT19937 RNG for 1:1 output parity with microgpt.py. Uses unsafe blocks for max performance. ~74x vs PyPy. |\n| [EEmicroGPT](https://github.com/Entrpi/eemicrogpt) | [@Entrpi](https://github.com/Entrpi) | C | \u003e440x | Single-file, dependency-free C. Explicit chain rule (no autograd overhead). SIMD/vectorized. Targets L1-resident weights on a single CPU core. |\n\n### Go\n\n| Name | Author | Language | Speedup vs CPython | Notes |\n|------|--------|----------|--------------------|-------|\n| [go-microgpt](https://github.com/prasad83/go-microgpt) | [@prasad83](https://github.com/prasad83) | Go | ~109x | Pure Go, no external dependencies. Clean modular layout: autograd, tokenizer, model, trainer. ~108.96x faster than original Python. |\n\n### JavaScript / TypeScript\n\n| Name | Author | Language | Notes |\n|------|--------|----------|-------|\n| [microgpt-ts](https://github.com/dubzdubz/microgpt-ts) | [@dubzdubz](https://github.com/dubzdubz) | TypeScript | Clean, fully typed code. Works in the browser. Zero dependencies. Live demo at [microgpt-ts.vercel.app](https://microgpt-ts.vercel.app/). |\n| [microgptjs](https://github.com/assassindesign/microgptjs) | [@assassindesign](https://github.com/assassindesign) | JavaScript (Node.js, ES5) | Node.js port. Supports training on poetry/Chinese text. |\n| [trainmyowngpt](https://github.com/jayyvk/trainmyowngpt) | [@jayyvk](https://github.com/jayyvk) | JavaScript | Browser-based visual UI. Real-time training visualization. Live demo at [trainmyowngpt.com](https://trainmyowngpt.com/). |\n\n### Julia\n\n| Name | Author | Language | Notes |\n|------|--------|----------|-------|\n| [microgpt_jl](https://github.com/ranton256/microgpt_jl) | [@ranton256](https://github.com/ranton256) | Julia (Flux.jl) | ~70x faster than Python (1.3s vs 89.2s). CUDA + Apple Metal support. Checkpoint persistence. 125 tests. Extended Shakespeare training mode. |\n\n### OCaml\n\n| Name | Author | Language | Notes |\n|------|--------|----------|-------|\n| [ocaml-microgpt](https://github.com/smimram/ocaml-microgpt) | [@smimram](https://github.com/smimram) | OCaml | ~6x faster than Python. Autograd library. |\n\n---\n\n## Python Variants \u0026 Extensions\n\n| Name | Author | Description |\n|------|--------|-------------|\n| [tapegpt.py](https://gist.github.com/mplekh/3afbdfb9f063cf531cfd3d00685cfdc0) | [@mplekh](https://github.com/mplekh) | Python with Wengert tape autograd (same algorithm as microgpt.cpp/rust-microgpt). 1:1 output parity with microgpt.py. ~4x speedup over original. Good for benchmarking algorithm vs compiler. |\n| [microgpt + EMA + ASCII plot](https://gist.github.com/santakd/a3645e3b58246e9b092a8633bf95659c) | [@santakd](https://github.com/santakd) | Adds argparse, rich metrics, EMA loss, ASCII loss curve (raw + EMA overlay), live progress bar with ETA. Still pure stdlib. |\n| [AttoGPT](https://gist.github.com/JGalego/26d617e5c939af0c32f3c16e4e392803) | [@JGalego](https://github.com/JGalego) | Golf'd microgpt compressed to ~50 lines. A readability crime scene. |\n| [microgpt (66 lines)](https://gist.github.com/karpathy/8627fe009c40f57531cb18360106ce95#gistcomment-6000490) | [@metacritical](https://github.com/metacritical) | Compressed to ~66 lines. Described as \"a compression experiment and a readability crime scene.\" |\n| [MoE microgpt](https://gist.github.com/credo92/4ba7d2db64bd6993864aaebbf13983cc) | [@credo92](https://github.com/credo92) | Adds a naive Mixture of Experts (MoE) layer to microgpt. |\n| [ko-microgpt](https://github.com/woduq1414/ko-microgpt) | [@woduq1414](https://github.com/woduq1414) | Korean-language microgpt. Live demo at [ko-microgpt.vercel.app](https://ko-microgpt.vercel.app/). |\n| [gpahal/microgpt](https://github.com/gpahal/microgpt) | [@gpahal](https://github.com/gpahal) | PyTorch-based micro GPT with RoPE, GQA, multi-GPU training, cosine LR, HellaSwag eval, 2-stage training with weight souping. |\n- [micro-attention-residuals](https://github.com/grahamannett/micro-attention-residuals) - A micro GPT implementation with Block Attention Residuals from MoonshotAI,\n extending microgpt with learned residual-memory mixing before each sublayer.\n\n---\n\n## Ecosystem \u0026 Extensions\n\n| Name | Author | Description |\n|------|--------|-------------|\n| [molequla](https://github.com/ariannamethod/molequla) | [@ariannamethod](https://github.com/ariannamethod) | Multi-organism autonomous training ecology. Go + C + Rust + JS + AML. Organisms grow architecture at runtime (ontogenesis), exchange \"DNA\" cross-training, reproduce via mitosis, self-regulate via SyntropyTracker. |\n| [purevlm](https://github.com/sailfish009/purevlm) | [@sailfish009](https://github.com/sailfish009) | Vision-Language Model in the spirit of microgpt. Most atomic VLM. |\n| [chuck-optimizer](https://github.com/ariannamethod/chuck-optimizer) | [@ariannamethod](https://github.com/ariannamethod) | Adam variant with self-awareness: self-modulates λ from loss trend, injects noise η to escape plateaus. Claims 38% faster convergence than Adam. |\n\n---\n\n## Benchmarks\n\nPerformance relative to CPython 3.14 (lower µs/sample = faster):\n\n| Implementation | µs/sample | Speedup vs CPython |\n|---|---|---|\n| CPython 3.14 | 713,200 | 1x |\n| PyPy 7.3.17 | 301,400 | 2.4x |\n| microgpt.cpp | 3,260 | ~219x |\n| rust-microgpt | 1,620 | ~440x |\n\n\u003e Go runtime: ~108.96x faster than original CPython (5:23.61 min vs 2.97s for 1,000 steps).\n\n---\n\n## Related Projects (Karpathy lineage)\n\n| Name | Description |\n|------|-------------|\n| [micrograd](https://github.com/karpathy/micrograd) | The autograd engine microgpt's `Value` class is based on |\n| [makemore](https://github.com/karpathy/makemore) | Character-level name generator; microgpt uses its names dataset |\n| [nanoGPT](https://github.com/karpathy/nanoGPT) | Full-scale PyTorch GPT training; the production counterpart |\n| [minGPT](https://github.com/karpathy/minGPT) | Earlier minimal PyTorch GPT re-implementation |\n\n---\n\n## Contributing\n\nPRs welcome! Please include: repo URL, author, language, brief description, and speedup/benchmark if available.\n\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Frupeshs%2Fawesome-microgpts","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Frupeshs%2Fawesome-microgpts","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Frupeshs%2Fawesome-microgpts/lists"}