{"slug": "ps5lm-plus-now-runs-ai-models-natively-on-a-jailbroken-ps5-s-gpu", "title": "PS5LM Plus now runs AI models natively on a jailbroken PS5's GPU", "summary": "A fork of Mert Cobanov's open-source PS5LM project, called PS5LM Plus, now runs as a native application on jailbroken PS5 consoles, freeing 11.4 gigabytes of memory versus roughly 6 gigabytes for the earlier payload version and routing inference through the console's AMD GPU instead of its CPU. On the GPU, Qwen 3.8 27B generates roughly 20 tokens per second versus about one token per minute on the CPU, while Meta's Llama 3.2 3B reaches around 124 tokens per second and Google's Gemma 4 E4B lands near 61 tokens per second. The setup requires a PS5 jailbroken on firmware between 7.00 and 13.60 and three payloads loaded in sequence (kstuff-lite, etaHEN and ShadowMountPlus), and Sony has never supported or acknowledged it.", "body_md": "*A new fork ditches the old payload workaround entirely, nearly doubling the memory available for AI models and letting them run through the PS5's GPU instead of its CPU.*\n\nPS5LM Plus, a fork of the open-source PS5LM project built by Mert Cobanov, now runs as a native PS5 application rather than a loaded payload. That single change matters more than it sounds. It frees up 11.4 gigabytes of memory instead of the roughly 6 gigabytes a plain payload could access. And it routes inference through the console's AMD GPU instead of its CPU. The jump from CPU to GPU is the real story here. On the CPU, Qwen 3.8 27B crawls at about one token per minute. On the GPU, the same model hits roughly 20 tokens per second.\n\nThat is not a typo-level difference. It is the difference between a demo and something you could actually sit and use.\n\nSmaller models move even faster. Meta's Llama 3.2 3B generates around 124 tokens per second. Google's Gemma 4 E4B lands near 61 tokens per second. Those numbers put a $500 console roughly in the same conversational speed range as a lot of budget desktop GPU setups running small quantized models. That's exactly why this keeps circulating among hobbyists rather than fading as a novelty.\n\nThe earlier version of PS5LM ran as a payload, a piece of unsigned code injected after the jailbreak chain finishes booting. Payloads on the PS5 are capped at a relatively small memory allocation, which is why the original project topped out near 6 gigabytes and stayed CPU-bound. PS5LM Plus instead installs as a proper application inside the jailbroken environment, which the console's own memory manager treats differently and allows to claim far more of the system's unified memory pool. The PS5's AMD GPU sits on 16 gigabytes of unified memory total, so 11.4 gigabytes represents the large majority of it, not a token increase.\n\n[Black Forest Labs Shrinks Flux 3 To Run On An 8GB Gaming GPU](https://startupfortune.com/black-forest-labs-shrinks-flux-3-to-run-on-an-8gb-gaming-gpu/)\n\nBlack Forest Labs priced its Klein API at $0.009 per image and licensed the weights free for companies under $5 million in revenue, undercutting the cloud GPU math that forced startups onto paid APIs. - [flux 3 runs on 8gb gaming gpu](https://startupfortune.com/black-forest-labs-shrinks-flux-3-to-run-on-an-8gb-gaming-gpu/) - [optimized image model for budget graphics cards](https://startupfortune.com/black-forest-labs-shrinks-flux-3-to-run-on-an-8gb-gaming-gpu/)\n\nThat extra headroom is what lets a 27-billion-parameter model fit at all. Qwen's 3.8 27B needs a GGUF quantization that lands comfortably under that ceiling. And llama.cpp, the inference engine PS5LM is built on, is exactly the kind of lightweight, CPU-and-GPU-flexible runtime that makes squeezing a large model onto unconventional hardware possible in the first place.\n\n## The catch: this requires real work and a real console\n\nNone of this runs on a PS5 fresh out of the box. Getting PS5LM Plus working requires a console already jailbroken on firmware between 7.00 and 13.60, plus three separate payloads loaded in sequence first: kstuff-lite, etaHEN, and ShadowMountPlus. That is not a one-click install, and Sony has never supported or acknowledged any of it. Jailbreaking a PS5 voids its warranty and can brick the console if a step goes wrong: the single biggest reason this stays a hobbyist curiosity rather than a consumer on-ramp to local AI.\n\nThere's also a real-world tradeoff once the model is running. Long prompts mean long waits. Prompt processing on the PS5's hardware is nowhere near as fast as the token generation numbers suggest, so feeding the model a big document or a long conversation history can leave you staring at a loading state for a while before generation even starts.\n\nStill, the direction of travel here is the part worth paying attention to. GPU and cloud inference costs dominate every AI headline this year. And here's a hobbyist project proving a five-year-old gaming console, something tens of millions of households already own, can run a 27-billion-parameter model fast enough to hold a conversation. It won't replace a cloud API for anyone running real workloads. But it's a genuinely useful data point in the argument that inference is getting cheaper and more decentralized faster than the infrastructure narrative suggests.\n\n**Also read:** [Google's Universal Cart Goes Live and Retailers Are Already Pushing Back](https://startupfortune.com/googles-universal-cart-goes-live-and-retailers-are-already-pushing-back/) • [Goolsbee says the AI building boom is now pushing inflation into unrelated industries](https://startupfortune.com/goolsbee-says-the-ai-building-boom-is-now-pushing-inflation-into-unrelated-industries/) • [Big Tech's worst week since the tariff crash exposes AI debt fears](https://startupfortune.com/big-techs-worst-week-since-the-tariff-crash-exposes-ai-debt-fears/)\n\n*This article is posted in [Technology News](https://startupfortune.com/category/technology/), check it out for more related stories.*\n\n## Join the discussion\n\n[Open in the community →](https://startupfortune.com/community/)\n\nAlmost there. Sign in and your reply posts straight away.", "url": "https://wpnews.pro/news/ps5lm-plus-now-runs-ai-models-natively-on-a-jailbroken-ps5-s-gpu", "canonical_source": "https://startupfortune.com/ps5lm-plus-now-runs-ai-models-natively-on-a-jailbroken-ps5s-gpu/", "published_at": "2026-10-17 01:15:17+00:00", "updated_at": "2026-10-17 01:59:08.067761+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-infrastructure", "ai-tools"], "entities": ["PS5LM Plus", "PS5LM", "Mert Cobanov", "Sony", "AMD", "Qwen 3.8 27B", "Llama 3.2 3B", "Gemma 4 E4B"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/ps5lm-plus-now-runs-ai-models-natively-on-a-jailbroken-ps5-s-gpu", "markdown": "https://wpnews.pro/news/ps5lm-plus-now-runs-ai-models-natively-on-a-jailbroken-ps5-s-gpu.md", "text": "https://wpnews.pro/news/ps5lm-plus-now-runs-ai-models-natively-on-a-jailbroken-ps5-s-gpu.txt", "jsonld": "https://wpnews.pro/news/ps5lm-plus-now-runs-ai-models-natively-on-a-jailbroken-ps5-s-gpu.jsonld"}}