Some people buy GPUs costing tens of thousands of euros, while others, to test the limits of local inference, dust off a handheld console launched twenty years ago. The experiment comes from a user on GitHub and shows a 90-million-parameter conversational LLM running on a Sony PSP, consumer hardware from 2004. The answer to the question “what can the PSP actually handle?” is clear: a 90M model is about the maximum it can manage before token generation becomes unbearable.

The numbers are those of a teaching test, not a service. The system produces between 0.5 and 0.6 tokens per second: a full reply can take from one to three minutes. This is not a missing optimization problem, but the reflection of a physical constraint. The PSP was not designed for transformer workloads, and its memory and CPU impose a very low ceiling on model size and inference speed.

The interesting point is not practical utility, which is close to zero. The model can write mediocre poems, short stories, non-functional code and, occasionally, answers simple questions correctly such as who makes MacBooks or what an LLM is. Other times it hallucinates nonsense answers. This is expected behavior for a minimum-size model: enough parameters to show some language ability, too few to be reliable.

The real lesson is structural. This experiment shifts attention from the data center to the extreme edge of the network, where hardware is old, memory is tight and no specialized GPU exists. It shows that local inference is not an exclusive territory of recent and expensive machines: with small enough models and some patience, obsolete devices can also run a self-hosted LLM. But it also shows that between “runs” and “is useful” there is a huge gap.

For those evaluating on-premise or edge deployments, the message is not “use a 2004 console”, but that the real bottleneck is almost always the relationship between available memory, model size and acceptable tokens per second. On AI-RADAR, the trade-offs between these constraints are analyzed in the section dedicated to on-premise LLMs.

The experiment will not change the market. No vendor will stop selling accelerators because of a PSP. But it is a useful reminder: the race toward ever-larger models coexists with an opposite movement, made of limit tests, compact models and hardware that nobody would still consider “compute”. In this sense, the PSP becomes an involuntary benchmark for those wondering how local AI can really be.