abhi245y/PS5LM-Plus

PS5LM Plus

llama.cpp on a jailbroken PlayStation 5, on its GPU. A native app that runs GGUF language models on the console itself, with a dashboard on the TV, a model store, and a chat and OpenAI-compatible API for every device on your network.

PS5LM Plus started as a fork of PS5LM by Mert Cobanov (cobanov/PS5LM): the llama.cpp port to the PS5, the payload and the early research come from there, and the project stays GPL-3.0 like the original. Since then it has grown separately around the native GPU app. The original payload's README is kept in docs/PAYLOAD.md.

What it does

A payload gets about 6 GiB of memory and no GPU. A native app (a title) gets the console's 12 GiB of direct memory and the GPU, so PS5LM Plus runs as one: llama-server with ggml's Vulkan backend on Mihawk-99's RADV port, in one 12 GiB device heap.

| Model (on the PS5 GPU) | Context | Generation | |---|---|---| | Qwen3.8-27B UD-Q2_K_XL (9.2 GiB) | 32k, q4_0 cache | 20.3 tok/s; a 7,775-token answer at 19.7 tok/s | | Qwen3.8-27B UD-IQ2_XXS (6.8 GiB) | 64k, q8_0 cache | 20.7 tok/s | | Gemma 4 E4B QAT q4_0 (4.8 GiB), from USB | 64k, q4_0 cache | 61 tok/s | | Llama 3.2 3B Instruct Q4_K_M (1.9 GiB), from USB | 64k, q8_0 cache | 124 tok/s |

For comparison, the 27B on the CPU as a payload, streamed from disk, made about one token a minute.

The model library at launch	Qwen 3.8 27B loaded: the plan and the arguments it runs with

Features

• Dashboard: generation speed with its history, GPU and CPU use, SoC power and energy per token, temperatures and fan, the memory pool and how the model, KV cache and the rest split it, context in use, storage on every drive, and the loaded model. Cross expands the USAGE, POWER and MODEL cards.
• Model library: every .gguf on internal storage, an M.2 drive or a USB drive, grouped by drive. Each shows the context and KV cache it would get and whether it fits; Cross loads it. A model on USB or M.2 loads straight from the drive.
• Settings: load a model at launch and which one, the longest context, the KV cache type, the chat's file tools and when to warn about their folder, sounds, and which drive downloads go to. Saved to /data/PS5LM/settings.json.
• Logs: app.log and llama.log on the TV, with an errors-only filter.
• Get models: Hugging Face's GGUF repositories, most downloaded or by family. Each file shows its size and whether it fits this console (Square checks one on demand, read from the first 16 MiB of the file). Cross downloads it.
• Downloads: a queue that survives a dropped connection, a pause and the app restarting. Each download resumes where it stopped, retries on its own, and shows speed and time left. Cross pauses or resumes, Square cancels. Files are checked against their SHA-256.

• Downloads check free space first (the file plus 1 GiB spare) and offer another drive when one has room.
• A download while a model holds the memory asks first, and can unload it for you.
• The chat's file tools are confined to /data/PS5LM/scratch, with a size warning and a Clear button.

_{The chat in a PC's browser, served by the PS5: Llama 3.2 3B on the console's GPU at 105 tok/s}

• Chat: llama.cpp's web UI at http://:8081, with file tools and, for vision models such as Gemma 4, images (put the model's mmproj file beside it).
• API: OpenAI Chat Completions and Responses, and Anthropic Messages, on the same port. opencode, codex, Claude Code and aider work against it: see docs/CLI.md.

• Planner: picks the longest context and the best KV cache type that fit the memory free, per model.
• Presets: sampling settings for Qwen 3.5/3.6/3.8, Gemma 4, gpt-oss, Granite, Nemotron, Mistral and Llama.
• Sensors and storage: a title may not read the temperature, fan or power sensors, or the real free space of a drive, so the app carries ps5-exporter and hands it to elfldr when it is not running.

What you need

A jailbroken PS5, with these payloads loaded before you start the app:

| Payload | Needed for | | |---|---|---| | kstuff-lite | the kernel patches homebrew titles run on | required | | etaHEN | homebrew support (see the note below) | required | | ShadowMountPlus | registers PS5LM Plus from /data/homebrew on the home screen | required | | elfldr (port 9021) | lets the app start ps5-exporter: thermals, fan, power and real free space | strongly recommended | | ftpsrv (port 2121) | installing and updating the app, copying models from a PC | recommended |

ps5-exporter comes inside the app; you do not load it yourself.

Install

1. Download ps5lm-plus-vX.Y.Z.zip from the releases and check it against its .sha256. 2. Unzip it and copy the PPSA99581 folder to /data/homebrew/PPSA99581 over FTP. ShadowMountPlus adds PS5LM Plus to the home screen. 3. Start PS5LM. Pick a model in Get models, or copy a .gguf to /data/PS5LM/models, to /data/homebrew/PPSA99581/models, or to PS5LM/models on a USB drive (exFAT) or the M.2 drive. 4. Choose the model in the library. Once the dashboard says Ready, open http://:8081 on a phone or computer.

To update, close the app and copy the new PPSA99581 folder over the old one. Settings, models and downloads stay in /data/PS5LM.

A title sees only its own folder unless something opens the console's storage to it. On the console PS5LM is developed on (13.00), ShadowMountPlus v1.7-beta5fix1 does so before the title starts, with or without a HEN: its debug.log shows [FAKELIB] directory mounted for PPSA99581: /data (and /mnt/usb0). OnionHEN does so on request: PS5LM asks when it starts (/download0/etahen_jailbreak), which OnionHEN honours for a title on its app jailbreak allowlist, so add PPSA99581 to app_jailbreak.exact_title_ids in /data/OnionHEN/config.ini.

When storage stays closed, the first start says so, with the firmware, and asks which way to go: get full storage (a current ShadowMountPlus, or the OnionHEN allowlist) and open the app again, or go on in the app's own space. Then settings, the log and downloads stay in the app's own space, /download0 (over FTP: /mnt/sandbox/PPSA99581_000/download0/PS5LM), whose size is downloadDataSize in sce_sys/param.json; models copied to /data/homebrew/PPSA99581/models load too. USB and M.2 drives stay out of reach. An update that raises downloadDataSize grows the space at the next start and keeps what is in it; lowering it frees nothing (tested on 13.00). Uninstalling the app deletes the models in both places, so update by copying the new folder over the old one.

If something goes wrong, open Settings > App > Debug report and press Cross (the first time, a popup says what the report holds before it is saved). The app saves one text file with its logs, settings, firmware, storage route, drives, models and which payloads are listening, and shows where it went: /data/PS5LM/debug/ over FTP, or /mnt/sandbox/PPSA99581_000/download0/PS5LM/debug/ when the app runs in its own space. Attach it to an issue. It holds no prompts or tokens, but it does include the end of ShadowMountPlus's log, which lists the games and images it mounts. Read it before posting, or paste it into an LLM and ask it to point out anything personal, and remove what you would rather not share.

Kernel log in reports (Settings > App, off by default) has the app collect the kernel log from the klogsrv payload (port 3232) from launch, keeping the last 1 MiB, and adds it to reports. klogsrv sends nothing from before a client connects, so turn it on, reproduce the problem, then save the report. The kernel log names other apps and services and can show network addresses and account IDs: turn it on only when asked for it; switching it on shows a popup that says so first. klogsrv sends every line to every client, so captures on a PC keep working while it is on.

Which models fit is in docs/MODELS.md; anything llama.cpp runs works if the library shows LOAD.

Controls

| Where | Button | Does | |---|---|---| | Everywhere | L1 / R1 | previous / next page | | Dashboard | D-pad, Cross | move between cards, expand one; Cross on MODEL opens the library | | Library | Cross | load the model under the cursor (the app restarts into it) | | Get models | L2 / R2 | change the search (Popular, Qwen, Gemma, ...) | | Get models | Square | check whether the file fits | | Get models | Cross | open a repository; on a file, download it | | Downloads | Cross | pause, resume, or try a failed one again | | Downloads | Square | cancel (press twice; the partial file is deleted), or clear a finished one | | Logs | Square, Triangle | switch log, errors only |

How it was built

The full notes are in docs/CONSOLE.md. In short:

• CPU first: Qwen 3.8 27B ran as a payload on the Zen 2 CPU, streamed from disk at about a token a minute: proof that llama.cpp works on the console, and that the GPU was needed.
• A title, not a payload: llama-server linked into an fSELF eboot.bin with a clean-room libc.prx, registered through ShadowMountPlus. Titles may not exit(), have some libc imports resolve to nothing, and need a working getaddrinfo, accept4 and HOME; each was found by a crash.
• The GPU: ggml-vulkan on PS5_Vulkan's RADV with no Vulkan loader, and RADV's unified heap for the whole 12 GiB in one heap.
• The screen: ps5-homebrew-ui through ps5-opengl, in the same process as RADV (the two Mesa builds share 4,399 symbol names, so the GL stack is linked with only its API visible).
• Switching models: unloading in place leaves about 5 GiB of the pool held, so the app restarts itself for each switch.
• HTTPS: downloads use the system's own libSceHttp2 and libSceSsl, so the app carries no TLS library or certificates.
• What a title cannot do: read sensors, see real free space (statvfs in a title is a stub), or start processes. ps5-exporter, a payload, covers the first two.

Building the app

The app links llama.cpp with ggml-vulkan against PS5_Vulkan's RADV build. Set up PS5_Vulkan and its forks in .deps/src (PS5_Vulkan, PS5_Mesa and PS5_PayloadSDK at the revisions PS5_Vulkan pins), then:

git clone --recursive https://github.com/abhi245y/PS5LM-Plus
cd PS5LM
(cd .deps/src/PS5_Vulkan && tools/setup-native-dependencies.sh && tools/build-radv.sh release && make app)
scripts/build-app.sh             # build/app-ps5/title/PPSA99581
scripts/release-app.sh           # build/release/ps5lm-plus-vX.Y.Z.zip and its .sha256

From the PC, with the app on the console:

scripts/preview-app-ui.sh                       # render every screen to build/app-preview/*.png (Mesa, no console)
PS5_HOST= scripts/ps5lm-app.sh cycle # close, deploy, launch; also close|deploy|launch
PS5_HOST= scripts/ps5lm-app.sh press r1        # press a button on the console
PS5_HOST= scripts/ps5lm-app.sh screenshot out.bmp

The app closes itself on request (/data/PS5LM/quit), so nothing is killed while it draws. Launching uses ps5-homebrew-dev-protocol's launch payload, cloned into .deps/src. /data/PS5LM/app-args.txt (one argument per line) overrides the planner.

| Path | What | |---|---| | ps5/app | the native app: dashboard, planner, store, downloads, llama-server on the GPU | | third_party/llama.cpp | upstream llama.cpp, pinned as a submodule | | third_party/ps5-homebrew-ui | the UI toolkit | | patches/ | PS5 changes to llama.cpp, kept small (tool confinement, Anthropic system messages, ...) | | ps5/compat, probes/ | the payload build's libc additions and memory probes (see docs/PAYLOAD.md) | | scripts/ | build, release, deploy and console tools | | docs/ | coding tools, models that fit, console notes, porting, research, plans |

Versions and branches

• Releases follow semver: tags vX.Y.Z, notes in CHANGELOG.md, version set in ps5/app/version.hpp and shown on the dashboard. v0.1.x are the original PS5LM payload releases, by Mert Cobanov.
• main is the integration branch. Each feature is built on its own branch (feature/..., docs/..., research/...) and merged into main.

Why llama.cpp

Earlier PS5 LLM work hand-writes GPU kernels for one model at a time. PS5LM ports llama.cpp itself, so every architecture and every quantization llama.cpp supports comes along: Qwen, Llama, Mistral, Gemma, from 1-bit to 8-bit.

Credits

• Mert Cobanov, who created PS5LM (cobanov/PS5LM): the llama.cpp port to the PS5, the payload and its library, the build scripts and the research PS5LM Plus started from.
• llama.cpp / ggml, which does the actual work.
• PS5_Vulkan (Mihawk-99) for RADV on the PS5, and ps5-homebrew-ui, ps5-opengl and ps5-homebrew-dev-protocol (BlackBearReloaded) for the app's screen and tooling.
• ps5-exporter (Marice) for the sensors and drive space the app shows.
• ps5-payload-dev (John Törnblom) for the SDK, the ELF loader, ftpsrv and klogsrv.
• PS5SX2, ProsperoAI and PS5_Vulkan, whose notes on the console's memory and GPU this project leans on.

Licence

GPL-3.0-or-later, see LICENSE. llama.cpp keeps its MIT licence. No Sony SDK files, keys, firmware or model weights are in this repository.

PS5LM Plus is not affiliated with Sony Interactive Entertainment or Alibaba. "PlayStation" and "PS5" are trademarks of Sony Interactive Entertainment.