Technical guide
Understanding the ChudGPT models
ChudGPT is a family of small custom language-model experiments. This guide separates model parameters, trained checkpoints, runtime context, and serving behavior so the versions can be compared without inflated claims.
Not affiliated with chudgpt.com, ChudGPT Next Gen, OpenChud, or other projects using the ChudGPT name.
Comparison
| Experience | Parameters | Checkpoint | Context | What changes |
|---|---|---|---|---|
| Public V20 | 20,999,184 | public_assistant_rebalance · training step 1400 | 1,024-token model context | Experimental school-subject, conversation, digital-literacy, Python, C#, Unity, and instruction-following curriculum |
| Public-Music V1 | 20,999,184 | Separate neural-only Music checkpoint | 1,024-token model context | Original full songs, partial lyrics, hooks, titles, style descriptions, mashups, and revisions |
| Code | 20,001,792 | Verified Plus checkpoint (live) | 4,096-token runtime context | Coding-only serving profile with up to 320 generated tokens and four-candidate selection |
| Pro | 20,001,792 | Plus checkpoint | 3,072-token runtime conversation context | Enhanced serving profile with four-candidate selection |
| Plus | 20,001,792 | Step 1400 | Model-specific runtime | Broader custom training and the raw checkpoint used by Pro |
| Ultimate | 14,016,384 | Step 1200 | 512 tokens | Custom checkpoint plus a local custom-only reliability layer |
| Buggy | About 14M | Current default | 512 tokens | Original deliberately unreliable experiment |
| Mega | About 13M | Early, intentionally poor checkpoint | 512 tokens | Independent model designed to be worse than Buggy and every other ChudGPT version |
Quick recommendations
Best overall / general useChudGPT-Public V20
Best for DiscordChudGPT-Public V20
Best public APIChudGPT-Public V20
Best for experimental original lyricsChudGPT-Public-Music V1
Best for deliberate chaosChudGPT Mega
Least reliable modelChudGPT Mega
Mega is intentionally worse than Buggy. It is the project's deliberate chaos experiment, not a dependable assistant.
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.
Experimental training results
Both candidates started from the active Music V1 weights. The curriculum-only run used 165 training and 27 validation examples for 140 steps, ending at response-only validation loss 4.5335. The mixed run used 4,383 training and 177 validation examples from the expanded corpus plus repeated genre curriculum for 300 steps, ending at 4.5775. Different splits make these losses unsuitable for deciding which writes better songs.
The active checkpoint and both candidates were evaluated on unseen prompts about a snowy night bus, a ferry departure, a movie theater reopening, and an escaped bee swarm. Both candidates failed full-song and short-song relevance checks; the curriculum-only candidate also produced broken raw English. Neither was promoted.
Use !chud music <prompt> in Discord or choose Music in Desktop. Existing Music links and separate session behavior are unchanged.
ChudGPT-Public V20
ChudGPT-Public V20 remains Astra's small experimental 20,999,184-parameter CUDA language model with a 1,024-token model context. The currently selected checkpoint is public_assistant_rebalance/latest.pt at training step 1400.
Its latest training prioritizes everyday assistant tasks, titles, basic facts, and simple code, with a small game-retention slice. Existing fallbacks and the separate LGBTQIA+ neural checkpoint are preserved. Topic-aware context selection now keeps genuine follow-ups while isolating explicit subject changes so stale replies are less likely to hijack the next request.
This is experimental curriculum expansion—not mastery, human-level intelligence, a specific human mental age, or guaranteed accuracy. Public can still produce incorrect, strange, funny, or unrelated output while preserving its intentionally odd CHUD personality. LGBTQIA+, emoji, and geography helpers cover narrow requests; game and modding replies remain generated. There is no new generic apology fallback or invented-model detector.
API base: https://chudgpt-public.vercel.app
ChudGPT Mega
Mega is an independently trained experimental model with approximately 13 million parameters. It uses an early, intentionally poor checkpoint and is designed to produce readable-but-broken, unpredictable, inaccurate, and occasionally nonsensical responses.
It is worse than Buggy and every other ChudGPT version by design. Expect incorrect facts, broken grammar, wrong arithmetic, abrupt topic changes, and gibberish. Mega exists for entertainment and observing early model-training behavior.
Never trust Mega for factual, technical, medical, legal, financial, or safety advice.
All 12 available models
Public V20 is the standard public conversational model. Music V1 is the separate experimental music model: lyrics are neural-generated, while a disclosed adviser analyzes supplied lyrics. The Main family contains Buggy, 700, 1300, 1500, 1600, Ultimate, Plus, Pro, Code, and Mega.
Every model has a keyless canonical API route. The desktop app exposes the same registry in its composer model picker. In Discord, use !chud models to list choices, !chud model to see the current choice, or !chud model <id> to select one; choices are stored per user and channel.
| Model ID | Published name | Family | Intended use | Canonical API |
|---|---|---|---|---|
public | ChudGPT-Public V20 | Public family | public_assistant_rebalance step-1400 experimental everyday assistance, writing, basic facts, code, and game-retention curriculum | Model info ↗ |
music | ChudGPT-Public-Music V1 | Public family | Separate experimental lyric checkpoint with neural song generation, rejection checks, and a disclosed supplied-lyric genre adviser | Model info ↗ |
buggy | Buggy ChudGPT | Main family | Intentionally serves an early broken checkpoint | Model info ↗ |
700 | ChudGPT 700 | Main family | Historical step-700 checkpoint experiment | Model info ↗ |
1300 | ChudGPT 1300 | Main family | Historical step-1300 checkpoint experiment | Model info ↗ |
1500 | ChudGPT 1500 | Main family | Historical step-1500 checkpoint experiment | Model info ↗ |
1600 | ChudGPT 1600 | Main family | Historical step-1600 checkpoint experiment | Model info ↗ |
ultimate | ChudGPT Ultimate | Main family | Older, more reliable 14M experiment | Model info ↗ |
plus | ChudGPT Plus | Main family | Conversational model and serving profile | Model info ↗ |
pro | ChudGPT Pro | Main family | Plus-based general profile with a larger runtime conversation window and stronger recovery | Model info ↗ |
code | ChudGPT Code | Main family | Programming-focused model profile | Model info ↗ |
mega | MEGA CHUD | Main family | Deliberately chaotic and least reliable experimental model | Model info ↗ |
How evaluation is interpreted
Fixed acceptance suite10 cases
Broader evaluationSeparate held-out suite
Rejected candidatesNot releases
The fixed suite checks required reasoning, constraints, C#/Unity, debugging, logic, creativity, personality, introductory GPU/CPU knowledge, unusual wording constraints, and a label puzzle. The broader held-out suite catches over-specialization. Passing a fixed suite does not prove complete general knowledge, and ChudLab does not guarantee a correct answer to every prompt.
Parameters are not context
A parameter is a learned numeric value inside a model. Parameter count describes model capacity; it does not describe how much conversation the serving system keeps in view. Code and Pro both use the 20,001,792-parameter Plus checkpoint, while their runtime contexts are 4,096 and 3,072 tokens respectively.
A checkpoint records training
A checkpoint is a saved set of learned weights from a particular training step. Routes 700, 1300, 1500, and 1600 expose historical snapshots for comparison. A larger step number is not a promise of better answers and is not the model’s parameter count.
Pro is a serving profile
Pro was not newly trained and is not larger than Plus. It evaluates four generation candidates, supports responses up to 160 generated tokens, and adds stronger recovery, topic handling, identity consistency, and code-block rendering on the existing CUDA backend.
Code is the dedicated coding profile
The live Code page uses the more dependable verified 20,001,792-parameter Plus checkpoint. Its specialized serving profile generates, explains, debugs, tests, refactors, and remixes code; supports follow-ups such as “explain it,” “fix it,” and “remix it”; renders fenced code blocks with Copy buttons; handles clarification and conversation context; uses audited coding helpers; and refuses unrelated non-programming conversations.
The separate coding candidate
A coding-only candidate was genuinely response-only supervised fine-tuned for 1,200 CUDA steps on 347 original project-authored coding conversations. No external datasets were used. The work added 300 new Code-specific tests and reached 427 total passing repository tests.
Passing automated tests did not make the candidate ready for promotion: unseen raw generations were still inconsistent and sometimes malformed. The candidate therefore remains an evaluated experiment rather than the checkpoint served by the live Code profile.
Supported coding areas
Code is strongest in Python, C#, Unity, Unity IMGUI, JavaScript, HTML/CSS, SQL, APIs, websites, applications, and game-development tools. It runs through the existing ChudGPT CUDA backend and still requires human review and real testing.
What the models are good for
These versions are intended for harmless experimentation with basic arithmetic, beginner code, simple conversation, and comparisons between checkpoints. Even the more reliable modes can produce incorrect facts or broken code.
What they are not
ChudGPT is not a frontier model and is not comparable to ChatGPT. The base models do not browse autonomously. Public's user-triggered search and page-summary tools are bounded serving features, not unrestricted browsing or guaranteed-current knowledge. No version should be treated as an authority or used for professional advice or safety-critical decisions.