Developer Runs 9B AI Model on PlayStation 4
An open-source project called Garlic LLM runs a 9-billion-parameter model fully offline on Sony's 2013 console.
Updated on Nov. 15, 2026 in Artificial Intelligence
Live Poll
Would you use an old console for AI tasks if it were possible?
Developer Leonardo Alegría released Garlic LLM, an open-source application that runs the Qwen3.5-9B model entirely offline on a PlayStation 4. The project does not require Linux or an external computer and achieves 31 to 39 tokens per second.
Why it matters
The project shows that a console never designed for machine learning can run a large language model locally, without cloud access. It points to how custom low-level GPU programming can extract useful AI performance from older consumer hardware.
The model fits into approximately 4.5 GiB of the console's 8GB shared GDDR5 memory and supports a 16K-token context window. Speculative decoding using Qwen3.5's Multi-Token Prediction head lifted generation from about 22 to 31-39 tokens per second, but the report does not measure output accuracy.
The players
Leonardo Alegría
Alegría is the developer who built Garlic LLM and wrote the custom GPU kernels that run the model on the PlayStation 4.
VideoCardz
VideoCardz is a technology news website that reported on the Garlic LLM project.
The details
Alegría wrote custom GPU kernels in OpenCL C, compiled them for AMD's GCN architecture, and communicated with the GPU through hand-written PM4 command buffers instead of a conventional graphics driver. The PlayStation 4's homebrew environment lacks a dedicated GPU compute API or machine-learning libraries, which made custom kernels necessary.
Timeline
• VideoCardz published its report on Garlic LLM on November 15, 2026.
Roadmap
The project builds on the PlayStation 4 homebrew scene, whose hardware lacks a native compute API or machine-learning libraries, so developers must write their own GPU code. It extends that tradition from running games and emulators to running a 9-billion-parameter language model.
Owners of the PlayStation 4 with homebrew access could in principle run a local AI assistant offline, avoiding cloud subscriptions and sending no data to external servers. The reported speeds and 16K context window indicate practical use for short tasks such as writing or code generation.
The takeaway
Hardware limits set by manufacturers do not fully determine what software can do, and determined developers can extend older devices in unexpected ways. Readers interested in local AI should weigh speed and accuracy trade-offs before relying on such setups.
Further reading
For related coverage, explore Artificial Intelligence news and analysis.
Source note: This article includes information reported by VideoCardz.
Live Poll
Would you use an old console for AI tasks if it were possible?
