The Console Under the TV Now Runs AI: Inside the PS5 Local Model Server
Synergised Consulting Ltd
Last Updated: 20 October 2026
The question "do I need special hardware to run AI privately?" now has an answer visible on a games console: a jailbroken PlayStation 5 runs language models natively on its own GPU at up to 207 tokens a second, and serves them to every device on the network through an OpenAI-compatible API. According to Tom's Hardware [3], jailbroken PS5s now run local models at over 100 tokens per second. The silicon for useful local inference is already in millions of buildings; what decides whether it is usable is setup, memory fit and workflow choice.
What PS5LM Actually Is
PS5LM is an open-source project that ports llama.cpp, the standard engine for running open-weight language models, onto a jailbroken PlayStation 5, so the console itself becomes a private model server: no PC running while it works, no cloud and no account. According to the project's own documentation [1], every layer of the model runs on the console's RDNA 2 GPU through the Vulkan backend, drawing on about 11.4 GB of the console's shared memory for the model and its context, and it answers on the same port in both OpenAI's and Anthropic's API formats, which is what lets ordinary coding and chat tools connect to it. A separate fork, PS5LM Plus, rebuilt it as a native application on the console, which unlocks more of the machine: according to the fork's own documentation [2], a native title gets the console's full 12 GiB of direct memory rather than the roughly 6 GiB a background payload receives.
From one token a minute to the GPU
The engineering story explains why the GPU matters more than the benchmark table. The project's first attempt ran a 27B model on the console's CPU, streamed from disk, and it produced roughly one token a minute [2], which is a demonstration rather than a tool. Moving the same class of model onto the GPU changed the economics entirely: the fork's native version sustains about 20 tokens a second on that same 27B model [2], and smaller models run far faster, at 207 tokens a second for a 0.8B model and 55 tokens a second for a 9B model with a 64k context [1]. The console's GPU has 448 GB/s of memory bandwidth, and the project measured the 9B model reading its weights at about 70% of it [1], which is the number that says this is a real inference machine, not a stunt. In discovery workshops, owners rarely arrive with a list of what could run where; the useful question is the one this project answers mechanically, which is what a specific device can actually hold and deliver.
What Changes With and Without It
The honest way to see what a local model server changes is a worked comparison, offered here as an illustration rather than a case study. Take a small business running a repetitive drafting task, such as classifying incoming supplier emails against a fixed list of categories. Without a local option, the work either stays manual or crosses to a cloud API, and the second path raises questions a founder may not have answers to: which data leaves the building, who holds the account, and what happens to the text. With a console-class server on the network, the same classification task runs against a model that never leaves the premises, addressed by whatever tool the team already uses, because the server speaks the standard OpenAI API format on port 8081 [1]. The trade is capability for control: a 9B model at 55 tokens a second will not match a frontier cloud model on hard reasoning, but for a bounded, repetitive task the gap is often irrelevant and the data locality is not. What actually gets decided in practice is which tasks sit in that category, and that list is built in conversation, task by task, rather than handed over by a vendor.
What It Is Not For
The limits are concrete, sourced and worth stating plainly, because they define who this is genuinely for. First, it requires a jailbreak: the project needs a PS5 on firmware 7.00 to 13.60 with homebrew payloads loaded [1], so a stock console on current firmware cannot do this, and nobody should jailbreak a device they are not prepared to take out of normal service. Second, the memory ceiling is real: about 11.4 GB of shared memory for model and context [1] means the largest practical model is a 27B heavily compressed to 2-bit quantisation at around 20 tokens a second [1][2], not a hundreds-of-billions frontier model, and it will not approach one on complex reasoning. Third, the honest label for this class of machine is capacity work rather than judgement work: the fit is long-running, bounded, slightly-below-frontier tasks, and a business that needs top-tier reasoning on sensitive text still has to make a different decision. This is also where the MSP pattern from our mastermind sessions is relevant: an owner with budget and appetite whose outsourced IT provider has no AI enablement in its service scope will not get to this kind of setup through that relationship at all, which is a governance gap, not a hardware one.
What This Signals for Hardware You Already Own
The narrow news is a clever port of llama.cpp to a games console. The broader signal is that the hardware floor for useful private inference keeps dropping, and the constraint that actually binds a small business is increasingly knowing what its existing devices can do, not buying new ones. According to the project's own measured tables [1], a single consumer console serves a 9B model at 55 tokens a second with a 64k context, which covers a meaningful share of the bounded tasks a small business runs daily. The practical takeaway for an owner in the first hundred days of owning a business is not "go jailbreak a PS5"; it is that a capability inventory of what you already own, measured rather than assumed, is a cheap first document, and it answers questions about cost, privacy and deployment before any purchase conversation starts.
