PS5LM, an Open Project, Turns a PlayStation 5 into a Local AI Server
jailbroken PS5
A PS5 changed to run software outside its normal setup.
GGUF
A file format for storing a language model.
decision model
A model that returns choices or scores without a long answer.
What happened
PS5LM is an open project that turns a jailbroken PlayStation 5 into a local AI server. It loads GGUF model files onto the console's GPU. A laptop, phone, or coding tool on the same network can send requests to it. The TV shows the model interface. The project says no other PC, cloud service, or account is needed while it runs.
The background
The project uses llama.cpp, software for running language models. Its Vulkan backend runs on the PS5's RDNA 2 GPU. The README says 11.4 GB of shared memory is available for the model and its context. The idea is not a new PS5-only model. It connects a common model format and runtime to a console's computing hardware.
PS5LM exposes interfaces compatible with OpenAI's ChatGPT API and Anthropic's Claude API. It also offers a browser page. Existing clients can point at the console. The project lists image-capable models and decision models. A decision model returns choices or scores without first generating a long answer.
Why it matters
This is more than chatting on a television. One console can serve several devices at home. The README presents PS5LM as a private model server. If its setup works as described, users can choose local processing instead of sending requests to a cloud service. The project also says it needs no account while running.
That promise comes with a high barrier. The listed setup requires a jailbroken PS5 on firmware 7.00 through 13.60. It is not a normal feature available to every console. Its usefulness depends on the console's memory, model size, and setup.
What is confirmed
The README reports these results on a PS5 Slim:
- Qwen 3.5 0.8B: 0.5 GB and 207 tokens per second.
- Qwen 3.5 9B: 5.3 GB and 55 tokens per second.
- Qwen 3.8 27B: 9.2 GB and about 20 tokens per second, with 64k context.
- d1 Omni 600M: three questions in 0.3 seconds; a 720p image and seven questions in 2.9 seconds.
A token is a small output unit. It is not always a whole word. These numbers come from the project's own table. The repository also says it is not affiliated with Sony Interactive Entertainment, the company behind PlayStation. It contains no Sony SDK files, keys, firmware, or model weights.
The post on Hacker News, a technology news site, reached 295 points and 74 comments. That shows community attention. It does not prove the software's performance or claims. The repository is the source for the technical facts; Hacker News is the source for the attention figure.
What remains unknown
The README does not provide an independent benchmark. We cannot tell from it how long sessions affect heat, power use, response time, stability, or answer quality. We also cannot assume every GGUF model performs equally. The project reports temperature and power screens, but those screens are not the same as an outside test.
The jailbreak requirement creates another limit. Firmware support may change. Future PS5 updates may alter compatibility. It is also unclear how much maintenance the project will receive. A successful demo is not the same as a dependable home server.
What to watch next
Useful follow-up evidence would be release notes, issue reports, independent speed and power tests, and results from more models. Watch whether the supported firmware range stays current. Watch whether several devices can use one console for long periods.
The bigger question is practical, not theatrical: can a game console provide stable local AI for ordinary users? PS5LM shows a compelling possibility. Its current page shows a project claim, not a broad industry result.
Sources: PS5LM repository and Hacker News post.
PS5LM: A Project That Makes a PS5 Run AI
📰 Full story: PS5LM, an Open Project, Turns a PlayStation 5 into a Local AI Server
A public project makes a prepared PS5 run AI for devices at home.
GPU
The part of a computer that performs many calculations.
GGUF
A file format for storing an AI model.
token
A small piece of text produced by a model.
💡 The gist
- PS5LM runs AI models on a prepared PS5.
- Phones and laptops can ask that PS5 questions.
- Hacker News attention is not proof of correctness.
PS5LM is an open project that uses a PlayStation 5 as a local AI server. It needs a jailbroken console. A jailbroken console has special software changes. We will not need the details here.
The project uses llama.cpp. This software runs language models. It loads GGUF files onto the PS5's GPU. GPU means the part that does many calculations. Other devices on the same home network can connect through a browser or familiar AI interfaces.
The reason this matters is control. The project says the console needs no cloud service or account while it runs. It presents the PS5 as a private model server. This may let a user keep work inside the home network. It also lets several devices share one loaded model.
PS5LM also lists decision models. They answer choices and scores without writing long text. It lists image input for some models too. These features show that the project is more than a toy chat screen. Some models can read images. The project also shows requests on the TV. This makes the console a shared home tool. The README does not compare it with a desktop GPU. It also does not compare answer quality with online models.
The README lists tests on a PS5 Slim. A small Qwen 3.5 model made 207 tokens each second. A larger Qwen 3.8 model made about 20 tokens each second. A token is a small piece of output. It is not always a whole word. Larger models also need more memory. These figures come from the project, not an independent lab.
The setup is not simple for most people. The README lists firmware 7.00 through 13.60. The PS5 must also be jailbroken. Model speed, heat, power use, and answer quality may change. The page does not prove that every GGUF model works equally well.
The post became a topic on Hacker News. It received 295 points and 74 comments. Those numbers show attention. They do not verify the project's claims. The next useful evidence would be outside tests and long-running reports.
Sources: PS5LM repository and Hacker News post.
💬 A PS5 can run AI, but it is not a universal replacement
The comments find the project exciting, but describe it as a trade-off: local models are possible, while speed and quality are limited.
- A jailbroken PS5 can run a GGUF model. One commenter self-reported 11.4 GB of shared memory, so it may be interesting for someone without a GPU.
- The example is Qwen 3.8 27B in 2-bit form. It uses less memory, but commenters say its answers are worse than 4-bit or full precision. Comparisons with Claude/Opus or GPT-4 are opinions, not benchmark results.
- Users self-reported long waits on hard problems and speeds of only tens of tokens per second. Another self-reported about 20 tokens/s, $0.70/day in power, and 2 million output tokens/day; these numbers were not independently checked.
- So it may fit limited jobs such as code inspection, server monitoring, footage, plant watering, or game NPCs, but not replace a top cloud model for everything. Some commenters also say the project description overstates its usefulness.
- Whether it is a better deal than an N150 mini PC, a DIY PC, an RTX 3060, or a BC-250 depends on local price and effort. People also debate distributed storage instead of GitHub, and the comments do not say whether games can run at the same time.
initial digest at 74 comments (revision 1). We fetched 74 comments and sampled 74 across the thread. These are HN users’ reports, not independently verified facts.
PS5LM: A Project for Making a PS5 Run AI
📰 Full story: PS5LM, an Open Project, Turns a PlayStation 5 into a Local AI Server
A game machine can make AI answers for a home.
PS5LM
A project that lets a PS5 run AI.
GPU
The part that does many sums.
Hacker News
A website where people discuss technology news.
PS5LM is a project that lets a PS5 run AI. It needs a PS5 with special changes. The PS5's GPU is a part that does many sums. It holds an AI model and makes answers. A phone or laptop can ask it questions at home. The project says it needs no faraway computer or account while running. Hacker News is a place where people discuss technology news. Its post got 295 points and 74 comments. Points show that many people noticed the post. They do not show that it is correct. The project lists small and big models. It says they run at different speeds on a PS5 Slim. This is the project's own report. Other people still need to test it. Not every PS5 can do this. It needs a special setup first. The idea is simple: a game machine can also run AI.
Source: PS5LM repository and Hacker News post.
💬 A PS5 can talk like a little AI, but it may be slow
It is a clever trick, but the PS5 is not suddenly a perfect AI computer.
- A jailbroken PS5 can run Qwen 3.8 27B. One person self-reported 11.4 GB of shared memory.
- The smaller 2-bit version saves space, but its answers may be worse. A user self-reported about 20 little pieces of text per second and about $0.70 a day for power; hard questions can still take a long time. The numbers and performance were not independently checked.
- It might watch things or help game characters talk. But another computer or board may be a better choice, and the comments still do not settle where to keep the code or whether you can play games while it runs.
initial digest at 74 comments (revision 1). We fetched 74 comments and sampled 74 across the thread. These are HN users’ reports, not independently verified facts.
💬 LLMs can run on a PS5, but practicality is conditional
The thread treats a jailbroken PS5 as an interesting local LLM box, but the comments disagree about its usefulness. The main caveats are 2-bit quantization, long inference waits, and hardware-price comparisons.
initial digest at 74 comments (revision 1). We fetched 74 comments and sampled 74 across the thread. These are HN users’ reports, not independently verified facts.