Jailbroken PS5 Actually Runs a 27B Large Model — Now the PS5 Is Going to Get More Expensive Again | Headline Version Final | 2026-10-17 | 1,226 Chinese Characters | Self-check Passed (No corner brackets / No numbering in H2 / Benchmark data in tables) A PS5 is being used as an AI server on a living room TV. No cloud, no internet, no cost — just open a web page and chat with a large model. Meanwhile

💡 What You Will Learn

Jailbroken PS5 Actually Runs a 27B Large Model — Now the PS5 Is Going to Get More Expensive Again | Headline Version Final | 2026-10-17 | 1,226 Chinese Characters | Self-check Passed (No corner bracke

📜 Table of Contents

• Fitting 27B into a game console by unlocking the GPU
• Media says "100 tokens per second," but 27B only hits 20
• The cost of 8x compression, and the real bottleneck nobody mentions
• But what's actually getting more expensive isn't new consoles
• At the end of the day, what's the value here

Jailbroken PS5 Somehow Runs a 27B Large Model — Now the PS5 Is Going to Get More Expensive Again

Headline version, final draft | 2026-10-17 | 1,226 Chinese characters | Self-check passed (no corner brackets / no numbering in H2 / benchmark data in tables)

A PS5 is being used as an AI server on someone's living room TV.

No cloud, no internet, no cost — just open a web page and chat with a large model. The DualSense controller is still there, and Spider-Man still runs — except now this machine's overtime hours are spent running a 27-billion-parameter model.

The person behind this is Reddit user AdIllustrious7246. They forked a project from two weeks ago into PS5LM Plus and pulled off something nobody had done before: unlocking the console's GPU so llama.cpp can actually use it.

Fitting 27B into a game console by unlocking the GPU

There's an AMD GPU sitting inside the PS5, but Sony has it locked down tight. Normally this GPU can only render graphics — you can't feed it any other instructions.

The breakthrough came from PS5_Vulkan — someone ported RADV, the open-source AMD GPU driver from Linux, to the PS5 and pried that lock open. That's the real technical core of the whole project.

The results were immediate. The old version stuffed llama.cpp into a system process as a payload, which only got it 6.25 GiB of memory — go over that and the main interface crashes. The new version takes the native app route, getting 11.44 GiB of device memory at once, with the GPU directly schedulable.

With nearly double the memory, there's finally room for the model.

Media says "100 tokens per second," but 27B only hits 20

This number needs to be untangled, because something got garbled in the retelling.

This is where that "over 100 per second" comes from. Look at the model name in the bottom right: Llama 3.2 3B, only 1.9 GiB — not even in the same league as 27B.

The claim of "over 100 tokens per second" comes from the 1.9 GiB Llama 3.2 3B — a model nearly ten times smaller than 27B. The actual 27B runs at around 20 token/s.

Some Chinese-language reports conflated the two numbers, even writing "20 per second" and "boosted to over 100 per second" in the same paragraph — contradicting themselves.

Incidentally, tokens aren't the same as Chinese characters, nor words. 20 token/s is roughly three to five Chinese characters of output per second.

The cost of 8x compression, and the real bottleneck nobody mentions

A full-precision 27B model needs 54 GB of memory; the PS5 has 16 GB of GDDR6 at most. That 8x gap is all brute-forced through quantization — the main config uses UD-IQ2_XXS, compressing each weight to about 2 bits.

The cost is plain to see. The developer admits in the docs that the harsher the quantization, the more error-prone the model and the weaker its reasoning. With 64K context enabled, memory usage shoots straight to 91% — no headroom left at all.

There's an even harder limit that 95% of reports don't mention: this machine's prompt processing speed is only 35–50 token/s. In other words, feed it a 20,000-token document and it'll grind for several minutes before spitting out the first character.

Fine for everyday chat, but if you want it to read through a long codebase and then edit code, you basically can't wait that long.

The image above is a screenshot of the developer's live panel while running 27B. 20.9 is the upper edge around the average speed, MIN is 14.9; the 50 tok/s in the PROMPT column is the prompt processing ceiling mentioned above.

But what's actually getting more expensive isn't new consoles

That "going to get more expensive again" in the title needs clarifying — whose price is going up.

First, the source of the price hike. AI data centers have been frantically hoarding memory capacity these past two years, and storage now accounts for over 35% of the PS5's total cost. On April 2 this year, Sony raised prices globally in sync: the US standard edition went from $549.99 to $649.99, the PS5 Pro to $899.99, and the China digital edition jumped a full 1,000 yuan.

This is cost inflation, and cost inflation ultimately has to be absorbed by someone.

So what does this jailbreak mean? It proves that the console's 16 GB of GDDR6 paired with 448 GB/s of bandwidth has real value for local AI inference. That figure is roughly four to five times that of an ordinary dual-channel DDR5 desktop — and what large-model generation speed is bottlenecked by is precisely memory bandwidth.

Once gamers worldwide realize the console gathering dust at home is a ready-made local AI server, what happens?

Machines that can run this need firmware held at 13.60 or earlier. That pool has a ceiling, and it won't grow.

So the "mining boom redux" that media widely fears is actually backwards. At 86W power draw and a 1.9 GiB small model, it's nowhere near data-center scale — no company is going to run commercial compute on PS5s. The real risk is something else — a batch of old-firmware jailbroken machines with a second use case, and their secondhand prices might start moving first.

At the end of the day, what's the value here

Players with a jailbroken machine can now hook their console into the LAN as an offline private AI backend. Phones and computers on the network can use it as a ready-made API. What you save is the monthly API bill; what you gain is data that never leaves home.

But for the vast majority of people, this is more a signal than a feature they can use right away.

What it really means is this: AI spent the past two years eating up all the memory capacity, forcing a company like Sony — which can't even change prices once a year — to break precedent in April; and now someone has turned around and squeezed the locked-down VRAM on a game console into compute.

The same batch of memory: on one side it's pushing console prices up, on the other it's being dug out by players to run large models.

This is probably the most surreal contrast of the year.