Experimental AI project
Meet ChudGPT.
Meet ChudGPT-Public V20 — Astra's small 20,999,184-parameter experimental CUDA language model, now serving the selected public_assistant_rebalance checkpoint at step 1400. Accuracy is not guaranteed; strange CHUD replies are still part of the experiment.
Experimental, not dependable. ChudGPT's answers are deliberately unreliable. Do not use them for facts, advice, or important decisions.
chudgpt — experimental output
ChudGPT
No. I am a tiny experimental model and may produce incorrect, strange, or inconsistent text. Treat my output as a demo, not a trustworthy answer.
Project identity
Independent, experimental, and built by Astra.
ChudGPT is an independent experimental AI project developed and maintained by Astra / ASTRA228b. The models documented on this site—including Public, Pro, Plus, Code, Ultimate, Buggy, Mega, and archived checkpoints—belong to this project's custom model ecosystem.
This project is not affiliated with chudgpt.com, ChudGPT Next Gen, OpenChud, or other unrelated projects using the ChudGPT name. Irony Hub AI is a separate, unrelated project and is not part of the ChudGPT experimental model family.
Official source repositories and releases are maintained through ASTRA228b's GitHub unless otherwise stated by Astra.
Meet the developer ↗Public APIs
Every model is callable.
No API key is required. Use one shared directory for accurate IDs and canonical routes.
Inference runs on owner-hosted CUDA systems. Each family has separate availability; one catalog being online does not mean the other is online.
Live interactive map
See the whole ChudGPT system.
Select every model, trace its connected APIs and apps, inspect live CUDA status and response latency, or explore how the Discord bot, desktop app, tunnels, capabilities, and operations fit together.
12 models2 API families live health polling
Open live system visual 
CHUDGPTLIVE MAPNew / Two minds. One world.
PUBLIC × BUGGY
Small models. Real choices.
Watch Public and Buggy move through a little world, meet neighbors, build memories, and make questionable life decisions. Talk to either character or change what they see and hear.
Each visitor gets a separate browser world. Decisions and dialogue use the live Vercel model APIs. This is an experiment in small models, not a simulation of consciousness.
Enter the simulation ↗
A ChudGPT App · Web + Windows
MultiMP3, now in your browser.
Use the quick Web edition for supported YouTube links, albums, MP3 quality choices, tracks, and ZIP exports—or choose Windows for persistent folders, history, recovery, local purchased-MP3 organization, and deeper settings.
01 / About
A tiny model experiment.
ChudGPT is a family of small, custom language-model experiments. The live Code profile uses the verified 20,001,792-parameter Plus checkpoint with specialized coding routing, context, rendering, clarification, and audited helpers.
A separate coding-only candidate was genuinely trained for 1,200 CUDA steps on 347 original project-authored conversations with no external datasets, but it was not promoted because unseen raw generations remained inconsistent and sometimes malformed. Plus, Pro, Ultimate, Buggy, and every archived checkpoint remain available and unchanged.
Not affiliated with chudgpt.com, ChudGPT Next Gen, OpenChud, or other projects using the ChudGPT name.
02 / What it is
Built for experimentation.
A transparent look at what exists and what it cannot promise.
MEGA
Intentionally worse than Buggy
Mega deliberately produces chaotic, inaccurate, partially readable gibberish. It invents words, breaks sentences, misunderstands questions, changes topics, and gives unreliable information.
Available now13M
Independent step-90 model
Mega is a separate 13,045,008-parameter model trained from random initialization on 2,400 original project-authored conversations and 529,136 training tokens—about 28× Mega v1—with no external datasets.
Available now</>
Dedicated coding profile
Code generates, explains, debugs, tests, refactors, and remixes code while refusing unrelated non-programming conversations.
Available now4096
Coding working context
Code uses a 4,096-token runtime context, supports up to 320 generated tokens, and evaluates four candidates. These are serving settings, not new training.
Available now347
Original coding dataset
A separate coding-only candidate was response-only supervised fine-tuned for 1,200 CUDA steps on 347 original project-authored coding conversations with no external datasets.
Available now427
Tested before promotion
The project added 300 new Code-specific tests and reached 427 total passing repository tests. The candidate was not promoted because unseen raw generations remained inconsistent and sometimes malformed.
Available nowPRO
Enhanced Pro serving profile
Pro evaluates four generation candidates and improves recovery, topic handling, identity consistency, and code-block rendering on the existing CUDA backend.
Available now3072
Larger runtime context
Pro keeps up to 3,072 tokens of conversation in view and supports responses up to 160 generated tokens. These are runtime settings, not new training.
Available now20M
ChudGPT Plus
Plus uses a separate 20,001,792-parameter model, expanded custom training data, and a local reliability layer.
Available nowU
Ultimate reliability layer
Ultimate combines a custom-trained checkpoint with a local custom-only layer for arithmetic, audited basic answers, and clarification.
Available now14M
Custom 14M model
The main demo runs a small custom-trained language model with roughly 14 million parameters.
Available now#
Archived checkpoints
Steps 700, 1300, 1500, and 1600 are saved moments from training.
Available now!
Deliberately unreliable
Outputs can be wrong, nonsensical, or inconsistent. This is part of the experiment.
Available nowUI
Browser chat
A simple interface sends a prompt to the backend and displays the generated text.
Available nowS
Session context
Conversation context is kept for the current browser session and can be cleared.
Available nowLAB
Learning project
Built to explore training, checkpoints, inference, and deployment—not production assistance.
Available now03 / Under the hood
From prompt to output.
A simple view of the experimental chat flow.
01
You write
Enter a harmless test prompt.
→02
Route selects
The current model or archived checkpoint is selected.
→03
Backend runs
The backend generates text with that checkpoint.
→04
Demo displays
The unreliable output appears in the chat.
Generated text may be false or nonsensical. Never treat it as authoritative.
04 / Models & checkpoints
Choose an experience.
Mega is the deliberately worst separate model; Code is the dedicated coding profile; Pro remains the strongest general serving profile. Every existing model and checkpoint remains available.
CurrentDefault Buggy
Buggy ChudGPT
The current custom 14M model demo intentionally serves an early broken checkpoint.
Open experimental demoUltimateReliable 14M version
ChudGPT Ultimate
The strongest custom ChudGPT version for basic answers, beginner code, arithmetic, and conversation. Still small, local, and experimental.
Try UltimatePlusProven 20M checkpoint
ChudGPT Plus
The new custom 20M model with broader training, stronger raw responses, computed arithmetic, and audited local answers.
Try PlusProStrongest general experience
ChudGPT Pro
The strongest current ChudGPT experience: the proven 20,001,792-parameter Plus checkpoint served with a 3,072-token runtime context, replies up to 160 generated tokens, four-candidate response selection, and stronger conversational recovery.
Try ProCodeDedicated coding profile
ChudGPT Code
The programming-focused ChudGPT experience for generating, explaining, debugging, and remixing code. It supports longer coding conversations, rendered code blocks, Unity and C# workflows, and focused software-development assistance.
Try ChudGPT Code13M · STEP 90 · WORSE THAN BUGGYMEGA CHUD
ChudGPT Mega / MEGA CHUD
A fully neural, separate 13M model served from its real step-90 checkpoint. Stronger anti-memorization filtering avoids recognizable training phrases, training-data commentary, and leaked instruction language, while slightly hotter, stupider sampling produces more original broken gibberish.
13,045,008 parametersReal step-90 checkpointFully neural generationStronger anti-memorizationAvoids training phrasesNo instruction leakageHotter samplingMore original gibberishNo keyword routingNo canned factual answersWorse than every ChudGPTExperimental entertainment
Trained from random initialization. No shared weights with Buggy or any other ChudGPT. Its 120 identity examples name it only ChudGPT.
Deliberately worse than every other ChudGPT. No keyword answer routing or canned factual answers—expect original nonsense, broken facts, incorrect math, and industrial-grade CHUD behavior. Never trust its facts, calculations, code, explanations, or advice.Try MEGA CHUDAll checkpoints
Archived training snapshots for historical comparison
Step 700
ChudGPT 700
An early saved training checkpoint. Expect poor and unpredictable output.
Open experimental demoStep 1300
ChudGPT 1300
A saved training checkpoint, not a different model size or quality tier.
Open experimental demoStep 1500
ChudGPT 1500
A historical snapshot from the model's training process.
Open experimental demoStep 1600
ChudGPT 1600
Another archived training snapshot with unreliable experimental responses.
Open experimental demo05 / Public model
Use ChudGPT-Public.
Try the independently trained public model, inspect its repository, or connect through its no-key API when the owner-hosted inference server is online.
V20 · RECOMMENDED PUBLIC
ChudGPT-Public V20
ChudGPT-Public V20 is Astra's small, experimental 20,999,184-parameter CUDA language model. It currently serves the public_assistant_rebalance checkpoint at training step 1400, rebalanced for everyday assistant tasks, basic facts, writing, and code.
Current checkpointpublic_assistant_rebalance · step 1400
20,999,184 parameters · 1,024-token model context.
The selected assistant rebalance completed 1,400 new training steps on 1,347 examples. Everyday tasks, basic facts, writing, and code take priority, with about 6% game-retention examples. All existing fallbacks and the LGBTQIA+ specialist are preserved. Public keeps its 21M architecture and can still get things wrong.
This does not demonstrate mastery or a particular human mental age. Game answers remain neural. Casual identity chat is neural and keeps the first usable draft; Music V1 keeps its separate checkpoint and endpoint.
Discord app12-model switching
Public Discord bot powered by ChudGPT-Public V20.
Public currently routes to public_assistant_rebalance at step 1400. Mention it or use !chud <message>. Use !chud models to list models, !chud model to show the current choice, !chud model <id> to switch, and !chud clear to clear the current conversation.
The base model cannot browse independently. The official bot may provide limited page or link context through a separate bounded web-reading feature.
!chud models
!chud model plus
!chud Explain recursion
!chud model publicDesktop 0.4.112-model desktop client
Use Public, Music, and the Main family.
The selector remains beside the message box and supports all 12 models. You can switch inside a visible conversation: the next message goes to the selected model while desktop history stays visible and saved locally. Private model memory does not transfer—each model keeps separate server-side session context.
WindowsmacOSLinux
12 keyless API models2 Public-family models10 Main-family modelsIndependent family health20,999,184 parameters · 21Mpublic_assistant_rebalance · step 1400Experimental CUDA language model1,024-token model contextTopic-aware conversation contextTitles + everyday assistant tasksBasic facts + simple codeGame knowledge retainedLGBTQIA+ + emoji awarenessState + world capitalsEveryday + multi-turn conversationDigital literacy
curl -X POST https://chudgpt-public.vercel.app/api/models/public/chat \
-H "Content-Type: application/json" \
-d '{"message":"Hello","session_id":"demo-user"}'Very small and experimental. Public can produce incorrect facts, invalid code, awkward grammar, or harmless nonsense. Verify important answers and never rely on it for high-stakes decisions.
MUSIC V1 · SEPARATE CHECKPOINTmusic: true
ChudGPT-Public-Music V1
October 13, 2026 · Expanded dataset and genre work. Music V1 remains a separate experimental 20,999,184-parameter checkpoint with a 1,024-token context and music: true API metadata. It attempts original songs, short songs, sections, hooks, titles, styles, revisions, and lyric mashups; reliable full-song quality remains unresolved.
The corpus now contains 3,453 conversations, up from 3,381. Its 72 additions represent 24 individually written original song groups as complete songs, supplied-lyric genre explanations, and isolated choruses. The focused genre curriculum grew from 8 groups / 32 conversations to 24 groups / 192 conversations, covering full and short songs, requested genres, choruses, production descriptions, and genre reasoning without prior chat context. No existing song lyrics or private user chats were used for these additions.
Song requests that include a title or genre now route to full-song generation. Correct Style: metadata is accepted, explicit genres influence candidate ranking, and section requests reject the wrong section. Short and full songs reject drafts missing the concrete subject; full songs also reject a contradictory requested genre.
Song lyrics remain neural-generated, with no canned lyric replacements. A separate supplied-lyric adviser provides a recognizable genre, instrumentation, and reasoning grounded in the supplied words. It is disclosed as music-genre-adviser in generation logs and does not write song lyrics.
Neither experimental training candidate was promoted. The active checkpoint remains unchanged. The service returns a generation failure when it cannot produce a relevant song. This tiny, funny, experimental model can still produce awkward grammar and musical slop; stronger lyric modeling or a larger human-reviewed corpus is needed before claiming consistently good full songs.
Verification reported for this update: 701 repository tests passed. The local Music API was unavailable and the public Music endpoint returned HTTP 502 when checked for the release report. These changes have not been verified live; this is a dated check, not a live service-status indicator.
20,999,184parameters1,024token context3,453corpus conversations24genre song groups192genre curriculum conversations701repository tests passed
Use it your way
Discord: !chud music <prompt>
Desktop 0.4.1: choose any of the 12 models beside the message box while keeping visible chat history stored locally.
Prompts go to the selected owner-hosted API. Music and Public keep separate server-side session context; model memory is not transferred.
Public training update · September 2026
Everyday tasks.
Still ChudGPT.
Rebalanced neural training for titles, basic facts, simple code, and everyday help. The small model still makes mistakes, but general requests should not turn into unrelated game descriptions.
Assistant tasks come first
The selected checkpoint trained on 1,347 examples, with about 6% reserved for game retention. Try a title, a basic explanation, or a short Python function. Unfamiliar wording and strict constraints can still fail.
Existing fallbacks preserved
Math, geography, identity, greetings, and emoji handlers keep their existing behavior. The focused LGBTQIA+ neural checkpoint is retained. General assistant and game replies continue to be generated.
Public retains LGBTQIA+ and emoji awareness, local geography answers, and its intentionally odd personality. Casual identity chat comes from the model, including its awkward mistakes. It can still invent facts and misunderstand requests. Music and the other family models keep their own checkpoints.
06 / Getting started
Explore responsibly.
- 1
Choose a checkpoint
Open the current model or a saved training step.
- 2
Try a harmless prompt
Experiment casually; do not seek facts or important advice.
- 3
Expect errors
Compare patterns and oddities without assuming anything is correct.
07 / Project facts
ProjectChudGPT
DeveloperAstra / ASTRA228b
Developer profilehttps://chudgpt-landing.vercel.app/astra
Public modelV20 · 20,999,184 parameters · decoder-only transformer
Public context1,024 model tokens
Public checkpointpublic_assistant_rebalance · training step 1400
Public curriculumGames · modding · model family · geography · awareness · replay
Public evaluationRaw checkpoint comparisons + live serving checks; experimental coverage only
Public update scope1,400-step assistant rebalance · existing fallbacks preserved
ModelCustom 14M language model
StatusExperimental demo
ReliabilityDeliberately unreliable
Mega modelSeparate · 13,045,008 parameters
Mega architecture8,192 vocabulary · width 288 · 11 layers · 6 heads · FFN 1,104
Mega context512 tokens
Mega training data2,400 project-authored conversations · no external or secondary datasets
Mega training tokens529,136 · approximately 28× Mega v1
Mega checkpointReal step 90 · validation loss 0.553259 · training perplexity 1.73
Mega identity120 examples name it only ChudGPT
Mega runtimeLocal CUDA
Pro modelSame 20,001,792-parameter Plus checkpoint
Pro runtime3,072-token context; up to 160 generated tokens
Pro selectionFour generation candidates
Code modelSame 20,001,792-parameter Plus checkpoint
Code runtime4,096-token context; up to 320 generated tokens
Code selectionFour generation candidates
Code scopeCoding only; existing CUDA backend
Code candidate data347 original project-authored conversations; no external datasets
Code candidate trainingResponse-only SFT · 1,200 CUDA steps
Code validation300 new Code tests · 427 total repository tests passing
Live Code checkpointVerified 20,001,792-parameter Plus checkpoint
Plus modelStep 1400 · 20,001,792 parameters
Ultimate checkpointStep 1200 · 14,016,384 parameters
Context512 tokens
PurposeTraining and checkpoint experiment
InterfaceBrowser-based chat
AffiliationNot affiliated with chudgpt.com, ChudGPT Next Gen, OpenChud, or other unrelated projects using the ChudGPT name. Irony Hub AI is a separate, unrelated project and is not part of this experimental model family.
Last updatedSeptember 2026
Experimental tester
Put every model on the lab bench.
ChudLab Beta is a transparent workspace for harmless model testing—not a reliability benchmark or production tool.
CHUDLAB BETA
One prompt. Twelve small models.
Chat with one model, run fresh side-by-side comparisons, inspect the exact API request, export visible results, and keep browser saving off unless you choose otherwise.
12-model directorySide-by-side comparisonRequest inspectorOptional local savingNo API keyNo automatic retries
Tester only. Live APIs may be slow, rate-limited, unavailable, inaccurate, or nonsensical. Never submit secrets or rely on the output.
08 / Links
Explore ChudGPT.
Open each model experience or read the detailed technical guide.
LABOpen ChudLab BetaTest prompts, compare two models, and inspect real API requests across all 12 ChudGPT models.↗MEGATry MEGA CHUDOpen the deliberately worst ChudGPT model. Never trust its output.↗COpen current demoLaunch the custom 14M experimental model.↗PROTry ChudGPT ProOpen the strongest current ChudGPT serving profile.↗20MTry ChudGPT PlusOpen the newest custom 20M version.↗</>Try ChudGPT CodeOpen the dedicated coding-only ChudGPT profile.↗UTry ChudGPT UltimateOpen the most reliable custom 14M version.↗GUIDEModel guideRead a detailed comparison of parameters, checkpoints, context, and serving profiles.↗APIPublic API guideUse all 12 keyless model APIs with canonical routes and copyable examples.↗DEVAstra / ASTRA228bCanonical profile for the developer and maintainer of this ChudGPT project.↗
09 / FAQ
Before you try it.
Do all ChudGPT models have public APIs?+
Yes. All 12 published models have keyless canonical info, chat, generate, and clear routes across two independently hosted families. Open the API guide for exact model IDs, schemas, examples, and availability notes.
Does ChudGPT Desktop run the models locally?+
No. Desktop 0.4.1 stores chats and preferences on your device, but sends prompts to the selected owner-hosted ChudGPT API. Its composer model picker includes all 12 models, and each model keeps separate server-side session context.
How do I switch models in Discord?+
Use !chud models to list IDs, !chud model to show your current choice, and !chud model <id> to select one. The bot keeps model choice per user and channel.
What is ChudGPT?+
An independent experiment built around a small custom-trained language model and a browser chat interface.
What is ChudGPT Mega (MEGA CHUD)?+
Mega is a completely separate 13,045,008-parameter model trained from random initialization. It does not share weights with Buggy or any other ChudGPT model. It uses a real step-90 checkpoint trained on 2,400 original project-authored conversations and 529,136 training tokens—about 28 times Mega v1—with no external or secondary datasets. Its 120 identity examples name it only ChudGPT.
Can I trust MEGA CHUD?+
Absolutely not. Mega is intentionally worse than Buggy. Expect partially readable gibberish, invented words, broken sentences, misunderstood questions, incorrect facts and math, abrupt topic changes, and unreliable code. It is experimental entertainment; never trust its facts, calculations, code, explanations, or advice.
What is ChudGPT Code?+
Code is a coding-only live profile built on the verified 20,001,792-parameter ChudGPT Plus checkpoint. It adds a 4,096-token runtime context, responses up to 320 generated tokens, four-candidate selection, coding-focused routing, context-aware follow-ups, fenced code blocks, Copy buttons, clarification handling, and audited coding helpers. It is not a frontier model and generated answers still require review.
Was a separate Code model trained?+
Yes. A coding-only candidate received response-only supervised fine-tuning for 1,200 CUDA steps on 347 original project-authored coding conversations with no external datasets. It passed the project’s automated evaluation, including 300 new Code-specific tests and 427 total repository tests, but it was not promoted because unseen raw generations were still inconsistent and sometimes malformed. The live Code page uses the more dependable verified Plus checkpoint.
Which coding areas does Code support?+
Its strongest supported areas include Python, C#, Unity, Unity IMGUI, JavaScript, HTML/CSS, SQL, APIs, websites, applications, and game-development tools. It can generate, explain, debug, test, refactor, and remix code, but its output still requires review and testing.
What is ChudGPT Pro?+
Pro is an enhanced serving profile built on the proven custom 20,001,792-parameter ChudGPT Plus checkpoint. It was not newly trained and is not a larger model. It adds a 3,072-token runtime conversation context, replies up to 160 generated tokens, four-candidate response selection, and stronger recovery, topic handling, identity consistency, and code-block rendering.
Is Pro comparable to ChatGPT or a frontier model?+
No. ChudGPT Pro remains a small, local, experimental language model. It has no live internet access and can still produce incorrect or incomplete answers.
What is ChudGPT Plus?+
Plus is a separate custom 20,001,792-parameter model trained from scratch and response-tuned on a larger custom dataset. It uses computed arithmetic and audited local answers for better reliability, but remains small, experimental, offline, and capable of mistakes.
Can I trust its answers?+
No. The app explicitly describes its answers as deliberately unreliable. Do not use it for facts, advice, safety, or important decisions.
What is ChudGPT Ultimate?+
Ultimate is the most reliable custom 14M ChudGPT version for basic arithmetic, beginner coding, simple factual questions, and normal conversation. It combines a custom-trained step-1200 checkpoint with a local custom-only reliability layer. It is still experimental, has no live internet access, and can be wrong.
What do 700, 1300, 1500, and 1600 mean?+
They are archived training checkpoint steps. They are not parameter counts, product tiers, or a promise that a higher number is better.
Does it remember conversations?+
It keeps context during the current browser session. Clearing the conversation resets that session history.
Is this ChatGPT or an OpenAI product?+
No. ChudGPT is independent and is not created, trained, endorsed, or officially supported by OpenAI.
What is it for?+
Treat it as a learning experiment and an entertaining way to compare a tiny model at different training checkpoints.
Support ChudGPT
Keep the experiment growing.
Help support the models and their little world. Choose what works for you, from free access to a contribution of up to $100.
Donate
Support the project through Cash App.
Make a donation ↗Simulator access
Sign in with Google and choose an access option. Free access includes 20 AI actions per day.
Get access ↗Already have access?
Return to Public and Buggy’s world.
Enter the simulation ↗Talk to Astra
For access and payment questions, add astra0181123 on Discord.
Manual supportPublic × Buggy / Two minds
Meet your next experiment.
Two small models. A shared world. Watch them live, talk to either character, and see what happens next.
Meet Public & Buggy Get simulator access