Llama.cpp on a jailbroken PS5: run any GGUF model natively on the console
github.com
October 16, 2026
5 min read
55/100
Summary
PS5LM is a homebrew application that runs GGUF language and decision models locally on jailbroken PlayStation 5 consoles using llama.cpp’s Vulkan backend and the console’s RDNA 2 GPU. It turns the PS5 into a private network model server without requiring a PC, cloud service or account while running. Devices on the local network can access OpenAI-compatible /v1/chat/completions and Anthropic-compatible /v1/messages endpoints at the console’s port 8081; Claude Code can use the Anthropic-compatible API. The application assigns up to 11.4 GB of shared PS5 memory to a model and its context, selects a context length and cache that fit available memory, and can download supported GGUF files from Hugging Face. Listed PS5 Slim performance includes 207 tokens per second for Qwen 3.5 0.8B Q4_K_M, 55 tokens per second for Qwen 3.5 9B Q4_K_M, and about 20 tokens per second with a 64,000-token context for Qwen 3.8 27B UD-Q2_K_XL. PS5LM also supports GGUF decision models with llama.cpp decision metadata, including d1, Clef and OpenJev models. These models serve TypeSafe’s /v1/systemone API for typed choices, scores and yes-or-no responses about text and images. PS5LM requires a PS5 on firmware 7.00 through 13.60 with a jailbreak and specified homebrew payloads, including kstuff-lite, etaHEN and ShadowMountPlus.
Key Takeaways
- PS5LM runs GGUF language and decision models natively on a jailbroken PS5 GPU through llama.cpp’s Vulkan backend.
- The console exposes OpenAI-compatible chat, Anthropic-compatible messages and TypeSafe-compatible decision APIs to devices on the local network.
- A Qwen 3.8 27B UD-Q2_K_XL model is listed at about 20 tokens per second and a 64,000-token context on a PS5 Slim.
- PS5LM supports models with image projectors, allowing compatible models to process images in chat and decision requests.
- Using PS5LM requires a jailbroken PS5 running firmware 7.00 to 13.60 and homebrew support payloads.
What the discussion said
Commenters treated the PS5 port less as a route to console gaming AI than as a provocative proof that a heavily locked consumer box can become a local-model machine. The novelty landed: several readers like the idea of repurposing idle hardware for continuous, private-ish inference, and one argued that a 27B Qwen variant at roughly 20 tokens per second is genuinely useful for forgiving background work such as codebase triage, monitoring, camera review, or plant automation. The claimed token-per-day economics also made the project look reasonable for someone who already owns the console and lacks a discrete GPU. The enthusiasm was sharply bounded by capability and access constraints. Skeptics argued that 11.4 GB of shared memory forces severe quantization, making the result an engineering feat rather than a compelling everyday assistant; others positioned it as both slower and weaker than current premium hosted models, even if it can approach older GPT-4-class open-model performance below 30B parameters. The jailbreak requirement further narrows the audience to expensive, old, unpatched consoles, while alternative used hardware was debated as a less troublesome inference option. Some readers looked ahead to LLM-driven NPCs and ever stranger devices hosting models, but that remained imaginative rather than evidence that the PS5 setup is ready for mainstream game AI. Overall, the room admired the hack more than it endorsed the platform as a practical local-AI target.
Where opinion split
The core dispute was whether constrained PS5 inference is useful work or merely a novelty demo. Supporters said low-bit 27B models can sustain respectable token generation for cheap, long-running, error-tolerant automation when the hardware is already sitting unused. Critics replied that the memory ceiling, harsh quantization, jailbreak friction, and gap from top hosted assistants make a normal GPU or cloud model the sensible choice.
Community Sentiment
Positives
- A jailbroken console turning into a local inference box is an appealing bit of hardware liberation, especially for people who already own one and have no GPU.
- Around 20 tokens per second from a 2-bit 27B Qwen model can still carry background jobs such as codebase investigation, monitoring, vision review, and irrigation control.
- The projected millions of daily output tokens make an idle PS5 a plausible always-on model host when imperfect answers are acceptable.
- Readers see local models spreading onto unexpected consumer devices, and that expanding hardware base could make lightweight AI deployment far less exotic.
Concerns
- The PS5's shared-memory limit forces aggressive low-bit quantization, so critics see the port as a clever technical demonstration rather than a satisfying general assistant.
- Even favorable comparisons place it well behind the strongest hosted models in speed and capability, limiting it to smaller open models and tolerant tasks.
- Jailbreaking requires old, unpatched hardware that has become costly and scarce, undercutting the idea that this is an accessible local-AI platform.
- At similar secondhand-hardware prices, commenters argue a conventional GPU offers a far easier path to useful inference with fewer compromises.
