PS5LM Runs a 27B Qwen Model on a Jailbroken PS5 at About 20 Tokens a Second
Software / news
PS5LM Runs a 27B Qwen Model on a Jailbroken PS5 at About 20 Tokens a Second
The GPL-3.0 app serves GGUF models over the home network, and its own README shows an 11.4 GB memory ceiling that decides which models fit.
A GitHub project called PS5LM turns a jailbroken PlayStation 5 into a home server for language models, and its author reports about 20 tokens a second from a 27-billion-parameter Qwen model on the PS5 Slim.
The repository belongs to a user named cobanov and grew out of an earlier native GPU app by abhi245y called "PS5LM Plus", according to its README. On the day it was read it showed 346 stars, 67 forks, 80 commits and 8 open issues. Stars measure curiosity, not how many people own a jailbroken console, a point the Terminal made about REA's 270,097 stars.
What PS5LM serves, and on which ports
PS5LM runs llama.cpp, the open-source C++ inference engine for GGUF model files, on the console's GPU. It exposes an OpenAI-style /v1/chat/completions endpoint and an Anthropic-style /v1/messages endpoint, plus a browser chat page on port 8081.
The task it fits is narrow and concrete: point an existing OpenAI or Anthropic client at http://<console>:8081 and have a box in the living room answer. The TV shows incoming requests, and a Console page charts temperature, power and load.
The app is licensed GPL-3.0-or-later. llama.cpp keeps its MIT licence, and the README says no Sony SDK files, keys, firmware or model weights are included. It also says PS5LM is not affiliated with Sony Interactive Entertainment.
The speeds are the author's, on one console
The README lists four models, all measured on a PS5 Slim. The Terminal did not run them and has no independent figure.
| Model | File size | Reported speed |
|---|---|---|
| Qwen 3.5 0.8B, Q4_K_M | 0.5 GB | 207 tokens/s |
| Qwen 3.5 9B, Q4_K_M | 5.3 GB | 55 tokens/s |
| Qwen 3.8 27B, UD-Q2_K_XL | 9.2 GB | about 20 tokens/s |
- Qwen 3.5 0.8B Q4_K_M207 tokens/s
- Qwen 3.5 9B Q4_K_M55 tokens/s
- Qwen 3.8 27B UD-Q2_K_XL20 tokens/s
Source: PS5LM README (author-reported), accessed 2026-10-17
The 27B row runs with a 64k context. Qwen's own model card lists Qwen3.8-27B under Apache-2.0 with a native 262,144-token context, so the PS5 build uses roughly a quarter of that window. The card lists 7,516,626 downloads last month for BF16 weights, and its recommended runtimes are Transformers, vLLM and SGLang. The 9.2 GB UD-Q2_K_XL file PS5LM runs is a different, far smaller build of the same model.
Why 11.4 GB decides which models fit
The README says the model and its context get 11.4 GB of the console's shared memory. The 27B file is 9.2 GB. That leaves 2.2 GB for the 64k context cache and everything else, by the Terminal's subtraction, not a figure the README states.
The 27B model only gets there at a 2-bit-class quantisation, the same trade Nvidia made with Nemotron 3.5 Bolt to fit 74B parameters into 24 GB. Quantising that hard costs answer quality, and the README publishes speed, not quality scores. The 9B model at Q4_K_M, at 5.3 GB and 55 tokens a second, is the one with headroom.
What you need before it starts
The prerequisites are heavier than the app. The console must be jailbroken and on firmware 7.00 to 13.60. The README lists three required payloads: kstuff-lite for kernel patches, etaHEN for homebrew support and ShadowMountPlus to put the app on the home screen. It recommends two more: elfldr on port 9021 for temperature and power readings, and ftpsrv on port 2121 for file transfer.
Install is a folder copy. Unzip ps5lm-app-vX.Y.Z.zip, copy the PPSA99575 folder to /data/homebrew over FTP, then drop a .gguf file in /data/PS5LM/models or let the app download one from Hugging Face. It also serves a 600-million-parameter decision model, d1 Omni 600M, through a /v1/systemone API: three questions in 0.3 seconds, or a 720p image plus seven questions in 2.9 seconds, per the README.
What the README does not say
The page does not name the exploit used to jailbreak the console, and it has no section on limitations or on the risks of modifying a console. Whether a firmware update from Sony would close the route is not addressed. Only the Slim is benchmarked.
The 8 open issues are the next place to look. The README points to the latest release zip as the supported route, and no release date appears on the page.
Sources
More in Software
- 01Sphinx, a Git Hook, Quizzes Developers on Agent-Written Code Before a Push Goes ThroughThe MIT-licensed 0.1 release sets a global Git hooks path, can block a push for 25 minutes and sends the diff to a reviewer model of the user's choice.
- 02Microsoft and CMU Start PRAISE, a Certificate for Early-Career Engineers Working With AIJunior Microsoft engineers take four CMU courses over two semesters with senior mentors, but neither announcement gives a cohort size, a cost or an eligibility rule.
- 03Lean 4.35 Adds Cross-Checking With a Claude-Written Kernel Proven ConsistentThe lake check command reruns a project through five kernels after an OpenAI model got four kernel soundness bugs to accept a proof of False this summer.
- 04CLI-Anything Reaches 54,400 Stars; GUI Agents Led CLI Agents 59.1% to 48.2% in One StudyThe HKUDS project generates command-line wrappers for desktop software, and the only comparison available shows its approach catching up only after skills were expanded.