TL;DR
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
PS5LM is a homebrew app that runs llama.cpp on a jailbroken PlayStation 5 and serves locally loaded GGUF models to devices on the same network. The project reports performance figures for selected models on a PS5 Slim, but those are developer-provided results, not independent benchmarks; it also requires compatible firmware and jailbreak payloads.
PS5LM, a homebrew project for jailbroken PlayStation 5 consoles, runs llama.cpp on the console’s GPU and serves locally loaded GGUF language models to other devices on a home network. The project says the console can provide an endpoint compatible with OpenAI and Anthropic client formats, making the PS5 a local model server rather than just a device for on-screen chat.
The project describes using llama.cpp’s Vulkan backend on the PS5’s RDNA 2 GPU, with up to 11.4 GB of shared memory available for a model and its context. Users can load GGUF files from internal storage, an M.2 drive or USB storage. A model planner reads each file and selects context and cache settings intended to fit the memory available, according to the project documentation.
PS5LM says clients on the same network can connect through OpenAI-compatible /v1/chat/completions and Anthropic-compatible /v1/messages endpoints, or use a browser chat page. It also describes a separate /v1/systemone endpoint for supported decision models, which return typed answers such as a choice or yes-or-no response rather than generating a long passage of text. The app can display requests and replies on a television-connected interface and chart console temperature, power and processor load.
For a PS5 Slim, the project lists Qwen 3.5 0.8B at 207 tokens per second, Qwen 3.5 9B at 55 tokens per second and Qwen 3.8 27B at about 20 tokens per second. The listed 27B configuration uses a 9.2 GB quantized model and a 64,000-token context. These are figures reported by PS5LM; the provided material does not describe an independent test method or results across other console models.
A Console as a Local Model Server
PS5LM repurposes a games console as a network-accessible local inference machine. If the project works as described on a user’s setup, a laptop or coding tool can send requests to a model running at home without sending those prompts to a hosted AI service. That could appeal to users who want to experiment with local inference or reuse available hardware.
The trade-off is that this is not a supported PlayStation feature. It depends on a jailbroken console and specific homebrew components, and the project’s own instructions restrict use to firmware versions 7.00 through 13.60. Readers should not treat the advertised speeds or compatibility as guarantees: model size, quantization, context length and available memory all affect practical performance.
As an affiliate, we earn on qualifying purchases.
Requirements Behind the Homebrew Port
The PS5LM instructions require a PS5 on firmware 7.00 to 13.60 that has been jailbroken. The listed setup also calls for payloads including kstuff-lite for kernel patches and etaHEN for homebrew support. ShadowMountPlus is used to place the app on the console’s home screen. Other tools for installation, copying files and reading system metrics are described as recommended rather than required.
The repository says the application is built with ps5-homebrew-ui and llama.cpp, using ggml-vulkan on a PS5 Vulkan/RADV stack. It credits contributors to the native GPU app it grew from, as well as the PS5 Vulkan and driver work. The project says it is not affiliated with Sony Interactive Entertainment and that its repository does not include Sony SDK files, keys, firmware or model weights.
““Your PS5 runs the language model. Every device on your network uses it.””
— PS5LM project description
As an affiliate, we earn on qualifying purchases.
Compatibility and Performance Still Vary
The supplied project material does not establish how reliably PS5LM runs across different PS5 hardware revisions, jailbreak configurations or model files. Although the project says llama.cpp-supported architectures and quantizations are available, that does not mean every GGUF model will fit or run well within the console’s available memory and compute limits.
The published speed figures are attributed to the project and focus on a PS5 Slim. No independent benchmark procedure, sustained-load results, comparison with a PC, or power and temperature measurements for those specific runs are included in the source material. It is also not clear how long-term use affects console stability, or what consequences modifying a console may have for support, services or warranty; users should check applicable terms and understand the risks before proceeding.
As an affiliate, we earn on qualifying purchases.
Project Releases and Testing
PS5LM directs users to its latest GitHub release for the app package and provides installation instructions, model suggestions and documentation for connecting coding tools. The repository also lists issues, a changelog and a roadmap, but the supplied material does not specify a dated next release or a planned performance test.
The next useful evidence would be independent testing across PS5 models and firmware setups, including sustained performance, stability and model compatibility. Until then, the project’s listed speeds and capabilities should be read as developer-reported results, and prospective users should verify that their console and jailbreak environment meet the stated requirements.
As an affiliate, we earn on qualifying purchases.
Key Questions
What does PS5LM do?
PS5LM runs llama.cpp on a jailbroken PS5 and can serve a loaded local GGUF model to other devices on the same network through compatible API endpoints or a browser chat page.
Does it work on an unmodified PS5?
No. The project’s instructions require a jailbroken PS5 on firmware 7.00 through 13.60, along with specified homebrew payloads.
Can it run any GGUF model?
The project says it supports architectures and quantizations recognized by llama.cpp, but not every model will fit or perform well. Memory use, model size, context length and compatibility matter.
How fast are the models?
PS5LM reports, on a PS5 Slim, 55 tokens per second for Qwen 3.5 9B and about 20 tokens per second for Qwen 3.8 27B in the listed configurations. These are project-reported figures; the source does not provide independent benchmark results.
Does using it send prompts to a cloud service?
PS5LM describes inference as running locally on the console and serving requests over a home network, with no cloud required by its stated setup. That describes the project’s design; users should check the configuration of any client they connect.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
