Recommended settings

Recommended settings

What to run, based on Test results on PS5 Linux and the console measurements on Benchmarks. Every PS5 differs a little; where a setting needs a test, it says so.

Linux on a PS5

Setting	Recommendation	Why
Distribution	Debian 13 with Mesa from trixie-backports, or Fedora, Ubuntu 26.04, Bazzite	The PS5's GPU needs Mesa 25.3 or newer; Debian 12 can't work. Debian 13 was 9% faster on prompts than SteamOS and gives the GPU 13.1 GiB instead of 7.5
How to run it	The TV service with no desktop (systemctl set-default multi-user.target)	The console experience, and the most memory for models
GPU memory	sudo tokenstation-memory set max (the packages do this on a PS5)	13.1 GiB: fits 14B models at Q4_K_M with 8k of context, or an 8B at 32k
Clock boost	Off (the default); disable ps5-linux-tools' ps5boost.service	+8% on prompts for every model, generation equal or better. Runs hotter; no harm seen
Fan	The stock curve	Full speed cools 5 C and changes no speed
GPU temperature limit	Leave the kernel's 85 C	90 C gained under 0.5%
Compute units	Test at 36 first (tokenstation-cu test 300), then try 40 and test again; stay at 36 if anything fails	On the test console 40 passed and gave +2 to +14% on prompts. 38 may equal 36 on your chip (probe tells you). Another chip may have a faulty unit
KV cache	f16 (the default)	The 8-bit cache is slower on RADV and fails the GPU check on Gemma
Context	8k by default; up to 32k for an 8B model if the memory allows	32k in f16 costs nothing in speed; it costs 4.5 GiB for an 8B
Helper model	Automatic (the default)	Helps only Qwen 3 on code; the setting turns it off where it doesn't pay
ts5-tune	Not needed on a PS5; run it on other GPUs	The built-in settings (wave32, 128-wide prompt split) were already the best on the PS5
Drive	The fastest you can boot from; NVMe if the loader allows it	Cold loads run at the drive's speed: 300 MB/s over USB SATA means 30 s for a 14B
Suspend	Don't	PS5 Linux doesn't support it

Which model

Want	Model	On a PS5 with Linux (tok/s, prompt / generation)
The best general chat, agent work, tools	Qwen 3 8B Q4_K_M	513 / 70
Code, fastest of the 7B class	Qwen 2.5 Coder 7B Q4_K_M	564 / 74
Plain chat without thinking	Llama 3.1 8B Q4_K_M	512 / 72
Fast answers, light on memory	Gemma 3 4B Q4_K_M (109 tok/s) or Llama 3.2 3B (146)	1068 / 109
A bigger model that still moves	Gemma 3 12B Q4_K_M	334 / 43
The most capable that fits	Qwen 3 14B, Qwen 2.5 Coder 14B, or DeepSeek R1 Distill 14B (thinking), all Q4_K_M	about 270 / 38
Spoken replies and R1 talk	Whisper Base (English) and the Kokoro Heart voice, downloaded from Settings	CPU

On the PS5 title (11.4 GiB), the same picks up to the 12B models; Gemma 3 12B keeps about 29k of context there thanks to its sliding-window layers. Stay at Q4_K_M: Q6_K and Q8_0 of an 8B are 25 to 40% slower to generate for a small quality gain. Prefer Gemma 3 12B over Gemma 4 12B for now (a little faster, the same memory).

For apps

• Serve with as much context as fits (the default). Coding tools send 10 to 15k tokens of instructions first.
• Tell the app the context_length that /v1/models reports, and leave temperature and the rest unset to get the settings checked on the PS5.
• Turn on Require a key and make one key per app. Never forward port 8081; use Tailscale.
• Qwen 3 for agents and tool use; add /no_think or enable_thinking: false when you want it quick. Leave tools out of plain chat with a Llama.

Engine variables worth knowing (ts5.env)

Variable	Default	Change it when
TS5_KV	f16	Never on Gemma; q8 only to squeeze a context in on a Qwen or Llama, at a speed cost
TS5_DRAFT_MODEL	automatic	You want a particular helper file, or off
TS5_GPU_MEMORY_GIB	what the driver reports	The driver's number is wrong on your machine
TS5_WEB_SEARCH	when a chat needs it	always, to offer the search tool on every chat
RADV_PERFTEST	cswave32 on Linux	Not on a PS5 (wave64 is 5 to 11% slower there); try on other GPUs via ts5-tune