{"rewrite":{"id":"r_fccdc80b1b3e868a51a45533","clusterId":"c_09297d3cc79022704eeab698","slug":"ps5lm-plus-runs-llama-cpp-models-on-a-jailbroken-ps5-gpu","model":"deepseek-v4-1-flash","headline":"PS5LM Plus Runs Llama.cpp Models On A Jailbroken PS5 GPU","summary":"A GitHub project called PS5LM Plus runs GGUF language models on a jailbroken PlayStation 5 and serves chat from a PC or phone on the same network. Developer abhi245y built it on PS5LM, Mert Cobanov's llama.cpp port. In the developer's measurements, Qwen3.8-27B at UD-IQ2_XXS generated 20.7 tokens per second, and Llama 3.2 3B Instruct Q4_K_M loaded from USB reached 124 tokens per second.","whyItMatters":"Running as a home screen app instead of a jailbreak payload lifts the usable memory ceiling from about 6GiB to 12GiB and opens the GPU through Vulkan, the difference between a demo and a model that can be left running.","webCardHtml":"\u003cp\u003eLaunching from the home screen rather than as a payload is the whole point of the build. The developer says a payload is capped at roughly 6GiB with no GPU access, while the app form uses 12GiB and the GPU through llama.cpp\u0026#39;s Vulkan support and a ported driver.\u003c/p\u003e\u003cp\u003eModel files load from internal storage, an added M.2 SSD, or a USB drive, and a search function pulls them from Hugging Face. The selection screen shows each model\u0026#39;s size and whether it fits in available memory, and context and cache limits adjust themselves. A dashboard on a connected TV reports token speed, GPU and CPU load, temperature, and memory use.\u003c/p\u003e","blueskyPost":"PS5LM Plus posts 124 tokens per second for Llama 3.2 3B from USB, but Qwen3.8-27B at UD-IQ2_XXS drops to 20.7. The quantization level, not the GPU, sets the ceiling.","twitterPost":"PS5LM Plus hits 124 tokens per second on Llama 3.2 3B but 20.7 on Qwen3.8-27B at UD-IQ2_XXS. Quantization, not the PS5 GPU, sets the ceiling.","threadsPost":"PS5LM Plus posts 124 tokens per second on Llama 3.2 3B from USB, while Qwen3.8-27B at UD-IQ2_XXS manages 20.7. The spread sits in quantization level, not the PlayStation 5 GPU. PS5LM Plus also serves chat over the network to a PC or phone.","newsletterBlurb":"PS5LM Plus, a GitHub project by abhi245y built on Mert Cobanov's llama.cpp port, runs GGUF models on a jailbroken PS5 and serves chat to a PC or phone over the local network. It launches from the home screen rather than as a payload, which the developer says raises usable memory to 12GiB and enables GPU computation through Vulkan. Measured speeds were 20.7 tokens per second on Qwen3.8-27B and 124 on Llama 3.2 3B Instruct.","attributionJson":"[{\"source\":\"GIGAZINE\",\"url\":\"https://gigazine.net/news/20261019-ps5lm-plus-gpu-ai/\",\"title\":\"PS5LM Plus: Running llama.cpp on a Jailbroken PlayStation 5 GPU\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":4759,"outputTokens":739,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1792401745,"createdAt":"2026-10-19T09:18:41.000Z","publishedAt":"2026-10-19T09:21:11.000Z","updatedAt":"2026-10-19T09:21:11.000Z"},"cluster":{"id":"c_09297d3cc79022704eeab698","canonicalTitle":"脱獄したPlayStation 5のGPUでllama.cppを実行できる「PS5LM Plus」","representativeArticleId":"a_e31bcac079ab9058a35ddc54","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[\"PS5LM Plus\"],\"studios\":[],\"people\":[],\"type\":\"news\",\"domain\":\"games\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-10-19T06:40:00.000Z","lastSeenAt":"2026-10-19T06:40:00.000Z","updatedAt":"2026-10-19T09:21:08.000Z"},"attribution":[{"source":"GIGAZINE","url":"https://gigazine.net/news/20261019-ps5lm-plus-gpu-ai/","title":"脱獄したPlayStation 5のGPUでllama.cppを実行できる「PS5LM Plus」"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":["PS5LM Plus"],"studios":[],"people":[],"type":"news","domain":"games","is_roundup":false},"keyFacts":null}