Caution
This is the ps5pro branch of an unofficial fork, tuned for the PS5 Pro
(86.9 -> 96.6 tokens/s with Mistral 7B Q4_0, plus Pro memory mode). It is not
made or supported by the ProsperoAI author. See
release/ps5pro/README.md for what changed; the
original project is https://github.com/blackbearreloaded/ProsperoAI.
Private, local generative AI for PlayStation 5 homebrew
Chat, create 512 × 512 images, generate short audio, and synthesize speech
with models running locally through the PS5 GPU.
The interface is built with ps5-homebrew-ui, fetched at a pinned commit when the app is built. This picture is the production frontend drawn on a PC with stand-in models. See the UI guide for the pages, how it is built, and what a run on a console showed.
Warning
ProsperoAI is an experimental project for validating generative-AI workloads on the PS5 GPU. It is not production software.
Note
vulkan-runtime uses one pinned Prism ggml/Vulkan runtime for GGUF text,
SD-Turbo images, Stable Audio and Kokoro voice on RADV. The dedicated AGC inference
backend is removed. CPU operations still handle remaining media kernels; RADV
continues to use the platform GPU driver. See current status,
media comparison and build instructions.
- Runs supported models locally without an account, cloud API, or conversation upload.
- Uses
vulkan-runtimefor text and media compute, including exact Ternary Bonsai 2 27B. - Searches and filters installed text, image, audio, and speech models in a scrolling library.
- Uses stable model IDs for selection and saved conversations, with no eight-model catalog limit.
- Offers Midnight and Daylight themes, a living backdrop, interface sounds, notices, reduced motion, high contrast, and larger text.
- Draws answers in Latin, Greek, Cyrillic, Chinese, Japanese and Korean with scalable faces.
- Stores independent text, image, audio, and speech sessions under
/data/prosperoai/sessions. - Supports DualSense navigation, right-stick conversation scrolling, the PS5 on-screen keyboard, and a physical USB keyboard.
- Ships without weights. Models offers verified presets and standalone GGUF downloads.
- Includes an HTTP server with a browser Workspace, model downloads and progress at
http://<PS5-IP>:11434/. - Connects OpenCode and other OpenAI-compatible clients to local PS5 text inference, with streaming responses and function-call support.
Important
ProsperoAI does not run on an unmodified retail console. It is intended for consoles you own with an already configured, compatible homebrew loader. This repository does not include an exploit, proprietary Sony SDK, firmware, system module, key, or model weight.
Important
Built on the PS5 Native App Boilerplate. ProsperoAI preserves its reproducible native build, packaging, deployment, and release foundation.
| Model | Purpose | Installed size | Hardware result |
|---|---|---|---|
| Mistral 7B Instruct v0.3 Q4_0 | Text to text | 4.35 GiB | About 1.4–1.5 s load and 25–29 tokens/s for short, resident-context replies |
| Qwen3.5 9B Q4_0 | Text to text | 6.37 GiB | 3.72 s switch/load; 1.46 s measured 44-token prefill and about 30 tokens/s decode |
| SD-Turbo FP16 | Text to image | 2.40 GiB | 512 × 512 image in about 76 seconds |
| Stable Audio Open Small FP16 | Text to audio | 1.32 GiB | 0.372-second stereo clip in about 62 seconds |
| Kokoro 82M FP16 | Text to speech | 165 MiB | 1.975 seconds of 24 kHz speech in about 79 seconds |
The linked repositories contain the exact directory layout, integrity hashes, upstream provenance, licenses, and preparation recipe for each model. Do not rename or mix their internal files.
The current release reads GGUF directly. These historical conversion tools are for
older releases and are not needed by vulkan-runtime. Install the equivalent GGUF
from Models before migrating a .ps5lm bundle. Known installed legacy Mistral/Qwen
IDs remain aliases to their GGUF models; the old bundle files are left untouched.
The open-source model preparation tools convert a supported
single-file GGUF into ProsperoAI's deterministic model.ps5lm,
tokenizer.ps5tok, and model.json bundle. Complete recipes for the validated
Mistral and Qwen models are included in the repository.
cd model-tools
./prepare-model \
--provenance recipes/mistral-7b-instruct-v0.3-q4-0.json \
/path/to/Mistral-7B-Instruct-v0.3.Q4_0.gguf \
--output-dir /path/to/PPSA99004The recipe verifies the exact upstream revision, byte count, SHA-256, tensor layout, tokenizer, and every generated output. See the model-tools guide for source-download commands, the Qwen example, recipe creation, supported GGUF types, and the steps required when adding a new architecture.
A recipe prepares weights for a backend already implemented by ProsperoAI. It cannot add an inference backend, tokenizer algorithm, or quantization format that the application does not support.
These are measured Alpha results from one PS5 on firmware 6.02, not guaranteed benchmarks. Text decode figures describe a loaded model with a short or reused context. The first prompt also pays model-loading and prefill cost. Attention work grows with conversation length, so long sessions respond more slowly. The 4,096-token text context boundary has been hardware-validated, but keeping active conversations below roughly 2,000 tokens is currently more comfortable.
Image, general-audio, and speech generation prioritize correctness over speed. Stable Audio currently emits only its short deterministic validation clip, and Kokoro still uses a correctness-first CPU fallback for part of waveform generation. Both are functional demonstrations rather than real-time paths.
- Download
PPSA99004.zipfrom the latest GitHub release. - Extract it. The archive contains a complete
PPSA99004/app folder and an intentionally emptyPPSA99004/models/directory. - Upload the complete
PPSA99004directory to/data/homebrew/. - Keep the console's local ELF loader (TCP port 9021) available. At launch ProsperoAI
asks PS5-Lapy-JB-Daemon for
access to the console's storage: a Lapy service that is already running is asked
first, otherwise the app starts the one-request helper it ships with (
lapy.elf). No separate payload upload is required. - Register or refresh the app folder in your homebrew loader, then launch ProsperoAI.
- Open Models and download a Text, Image, Audio or Voice preset. Alternatively,
copy standalone text GGUF files or complete curated model folders to
/data/prosperoai/models/.
The browser interface is at http://<PS5-IP>:11434/. Both interfaces show verified
preset downloads and progress. Text and media inference use vulkan-runtime; text models use GGUF.
Prepared legacy .ps5lm bundles require their equivalent GGUF for migration.
Everything the app keeps is under /data/prosperoai/:
/data/
├── homebrew/PPSA99004/ the app folder
└── prosperoai/
├── models/
│ ├── Mistral-7B-Instruct-v0.3.Q4_0.gguf
│ ├── kokoro-82m-fp16/
│ └── sd-turbo-fp16/
├── sessions/ conversations, generated images and audio
├── logs/ the debug log, when it is switched on
└── prosperoai.cfg settings
Keep each downloaded model folder intact. ProsperoAI discovers all valid model folders at launch and shows their friendly names and purposes in Models. If no compatible model is installed, the app opens normally and explains where to add one.
Coming from 01.001.000: models in /data/homebrew/prosperoai/models/ are moved to
/data/prosperoai/models/ the first time this version starts, and settings and
conversations are copied over once. The storage helper of that version is no longer
used; if it is still running it is told to stop.
Without storage access (no ELF loader and no Lapy service) the app still opens and
keeps its data in its own sandboxed folder; models copied to /data/prosperoai/models/
are not visible to it then.
Only the app-folder ZIP is distributed; it contains no model weights.
Settings > Diagnostics > Debug log makes the app write a timed trace of what it does:
the models it found, each request with its result and figures, every message it showed,
downloads with their speed, and what the model runtimes themselves report. It is off by
default. Switch it on, repeat what went wrong, close the app and send us
/data/prosperoai/logs/debug-trace.txt (the run before it is kept beside it as
debug-trace.prev.txt) in a
GitHub issue. The log holds
your prompts and the answers of that run; read it before you share it.
Once per launch ProsperoAI asks the homebrew.page catalog whether a newer version is listed, and verifies the catalog's signature before it believes the answer. When there is one, it says so:
- Update now downloads the release from GitHub, checks it against the SHA-256 the catalog lists and closes the app. A small helper then puts the new version in place, and the console says when it is done: open ProsperoAI again.
- What's new shows the release notes first.
- Skip leaves everything as it is; the question comes back at the next launch.
Nothing is changed before the download has been checked, and a cancelled or failed update
leaves the app as it was. Models, settings and conversations are not touched: they live in
/data/prosperoai. Like storage access, the update needs the console's local ELF loader
(TCP port 9021).
This check and the downloads you start are the only requests the app makes on its own: one to homebrew.page (and to the catalog's mirror on GitHub Pages when that does not answer), naming the app and its version.
The HTTP server starts automatically when ProsperoAI launches. Keep the app running
on the PS5 and connect your computer or phone to the same network. Replace
<PS5-IP> below with your console's local IP address.
- Open
http://<PS5-IP>:11434/in your browser. - Use Models to download a preset or select an installed model. Download progress is shown while the files are fetched and verified.
- Open Workspace and choose Converse, Imagine, Compose or Speak. A card selects an installed model for that purpose, or opens its Models category.
- Use Back to Workspace to return; conversations can be deleted from their list.
Closing the PS5 app also stops its HTTP server. The API uses plain HTTP and requires no key by default; use it on a trusted local network. Optional bearer authentication and its effect on the browser interface are explained in the API guide.
ProsperoAI exposes an OpenAI-compatible API at http://<PS5-IP>:11434/v1, including
streaming text and function calls. OpenCode runs on your computer and uses the PS5
for model inference. First install a text model and check its exact ID:
curl 'http://<PS5-IP>:11434/v1/models'Add this provider to your project's opencode.json. Replace <PS5-IP> and, if
needed, the model ID with the value returned above:
{
"$schema": "https://opencode.ai/config.json",
"provider": {
"prosperoai": {
"npm": "@ai-sdk/openai-compatible",
"name": "PS5 ProsperoAI",
"options": { "baseURL": "http://<PS5-IP>:11434/v1" },
"models": {
"Mistral-7B-Instruct-v0.3.Q4_0.gguf": {
"name": "Mistral PS5",
"limit": { "context": 4096, "output": 512 },
"tool_call": true
}
}
}
},
"model": "prosperoai/Mistral-7B-Instruct-v0.3.Q4_0.gguf"
}Run opencode in that project, or select the model with /models. This follows
OpenCode's custom provider configuration.
Streaming and a tool-call roundtrip have been verified on PS5. Automatic tool
selection depends on the model; the tested Mistral model is not yet reliable for
autonomous coding tasks. See OpenCode setup and API limits for
larger contexts, authentication and detailed compatibility notes.
| Input | Action |
|---|---|
| D-pad / left stick | Move through models, conversations, or settings; Left reaches the conversation list |
| Cross | Start or open a conversation, write or send a prompt, choose a model, change a setting |
| Circle | Close a dialog, leave the conversation list, or return to the settings categories |
| L1 / R1 | Switch between Workspace, Models, and Settings |
| Square | Start a new conversation in Workspace; search in Models |
| L2 / R2 in Models | Previous / next model category |
| Triangle | Insert Space while typing; otherwise retry, delete a focused conversation, or play saved audio; in Models, cancel the running download |
| R2 while typing | Done; send from Workspace |
| Options | Ask to close the app, once nothing is running and the conversation is saved |
| Right stick | Scroll through the current conversation |
| Physical USB keyboard | Type in the prompt field or the model search; Enter sends; Tab is R1 and Shift+Tab is L1 |
The row at the bottom right of every page names what the buttons do there.
Each conversation is an independent session. Text and metadata are saved under
/data/prosperoai/sessions/; generated image and audio files live inside
the matching session directory. Deleting a session removes its associated
content. ProsperoAI does not send these files to a network service.
Ideas we would like to work on. They are not promises and carry no dates or order; open an issue to argue for one or to suggest another.
Models and generation
- Faster sound generation: a Stable Audio clip takes several minutes in the combined build.
- Vision models: let a text model look at a picture.
- Image-to-image, and variations of a generated picture.
- More image models beside SD-Turbo, with size and step settings.
- Speech to text: dictate a prompt through the DualSense microphone.
- Spoken answers: read a text answer aloud, with a choice of voices.
- Personas and system prompts, saved per conversation.
- Generation settings remembered per model.
Model library
- Resume an interrupted download instead of starting again, and clean up what a closed app left behind.
- Delete a model from the app, with the space it frees.
- A download queue.
- A Hugging Face token for repositories that need an account.
- GGUF files above 7 GiB, and split files.
- Before downloading a model, a hint of the memory it needs and the speed to expect.
Conversations
- Longer conversations and answers: 64 messages of 4 KB each are kept today.
- Export a conversation or a picture to USB, or fetch it from the browser page.
- Search, rename and pin conversations.
- Documents as context: ask about a text or PDF file from USB.
Console experience
- Menus in the console's language.
- A phone as keyboard and remote.
- Work in the background, with a notice when a picture or a download is done.
- Remember the last model whatever the length of its name.
Server and integrations
- Pictures, speech and sound over the OpenAI-compatible API, beside chat.
- An embeddings endpoint.
- A server page in Settings: address, key, connected clients, and a switch to turn it off.
Under the hood
- A faster update download.
- A cleaner debug log.
- An in-app benchmark: tokens per second and time per picture for each model.
Build from Linux, WSL, or an Ubuntu-compatible CI runner:
sudo apt update
sudo apt install ccache clang-18 clang-format-18 clang-tidy-18 lld-18 make \
ninja-build pkg-config python3 python3-venv tar unzip wget
make lint test
# Set up Vulkan dependencies as documented in docs/BUILDING.md.
make app-releaseOutputs:
dist/PPSA99004/ complete model-free app folder
dist/PPSA99004.zip archived model-free app folder
The build downloads and verifies the public PS5 Payload SDK, the ps5-opengl SDK, the
ps5-homebrew-ui interface kit, the Noto Sans faces for Chinese, Japanese and Korean, and zlib
inside the ignored .deps/ directory. It bakes the interface fonts, rebuilds the clean-room
libc.prx runtime, compiles the native app through parallel incremental Ninja builds,
caches compiler results with ccache, signs the executable, validates assets,
and assembles the release. Set USE_CCACHE=0 to disable the cache or
BUILD_JOBS=<count> to limit parallel compilation. Model weights are never
downloaded by the app build.
The Build workflow runs on pull requests,
version tags (NN.NNN.NNN), and manual dispatch. A push to main builds
nothing; a build on main is started by hand (Run workflow). It:
- restores dependency and ccache data;
- validates source, metadata, presentation assets, and the executable writer;
- reproduces and verifies the clean-room runtime shim;
- builds pinned RADV and llama.cpp inputs, then the vulkan-runtime model-free folder and
PPSA99004.zipwith Ninja; - rejects an artifact containing model data;
- writes
SHA256SUMSand uploads both release files; and - publishes those verified files when triggered by a version tag.
A release is made by pushing the version tag: the workflow builds, attests and publishes
PPSA99004.zip and SHA256SUMS; release files are not attached by hand. If no release exists
for the tag, the workflow creates it with the notes from CHANGELOG.md. If one exists without
a ZIP (notes written in advance, or a draft), it adds the two files and leaves the title and
notes alone. If one already has a ZIP, nothing is replaced and the run ends with a warning.
A release ZIP built by the workflow can be checked with
gh attestation verify PPSA99004.zip -R blackbearreloaded/ProsperoAI (GitHub CLI); this covers
releases built by GitHub Actions from now on, not earlier ones.
Every pull request gets an installable build named by its number and commit: see Pull-request builds.
src/ ProsperoAI UI, sessions, model routing, and GPU runtimes
vulkan/ Shared GGUF/media inference and RADV interface
include/ Application interfaces
assets/ The last-resort bitmap font (other fonts and sounds come from the kit)
ui-kit/ Which sources of the fetched interface kit are compiled
host/ Deterministic native-frontend capture harness
tests/console/ Scripts the app plays by itself in a test run on a console
sce_sys/ PS5 title metadata, artwork, icon, and selection music
models/README.txt Model-free release placeholder and install guidance
model-tools/ GGUF converter, validated recipes, and porting guide
vendor/ Pinned headers and static runtime dependencies
runtime/ Clean-room libc runtime inputs and integrity record
tooling/native/ Native ELF/FSELF and runtime build tools
tools/ Build, dependency, validation, packaging, and deploy scripts
| Identity | Value |
|---|---|
| Shell title | ProsperoAI |
| Title ID | PPSA99004 |
| Current app version | 01.004.000 |
| Writable data | /data/prosperoai (the app's sandbox without storage access) |
| Compute backend | vulkan-runtime on RADV, with documented CPU media operations |
Built with the PS5 Payload SDK by John Törnblom (ps5-payload-dev). Third-party components, authors and licenses are listed in THIRD_PARTY_NOTICES.md.
Copyright © 2026 BlackBearReloaded. Licensed under GPL-3.0-or-later; see LICENSE. Third-party components keep their own licenses.
- No affiliation. This is an independent homebrew project. It is not affiliated with, endorsed by, or sponsored by Sony Interactive Entertainment. "PlayStation", "PS5" and related marks are trademarks of Sony Interactive Entertainment Inc. Model names belong to their authors, who do not endorse this project.
- No proprietary material. No Sony SDK, firmware, encryption keys or decrypted system modules are included.
- No warranty. This project is provided "as is", without warranty of any kind, to the extent permitted by law. See sections 15 and 16 of the GPL.
- Use at your own risk. Running homebrew requires a modified console, which may void its warranty, breach the platform's terms of service, or cause data loss.
- Legal use only. Use it only with hardware, accounts and content you own. This project does not support or enable piracy.
This project was developed with AI assistance from OpenAI and/or Anthropic tools.
