What happened
The PS5LM Plus repository describes a native PS5 application built around llama-server, llama.cpp and ggml’s Vulkan backend running on a RADV port. The project began as a fork of Mert Cobanov’s PS5LM; the README credits that project with the original llama.cpp port, payload and early research. PS5LM Plus keeps the GPL-3.0-or-later license and has grown into a GPU-enabled app rather than only a payload-based proof of concept.
The README says a payload gets about 6 GiB of memory and no GPU, while a native title gets 12 GiB of direct memory and GPU access. It reports Qwen3.8-27B UD-Q2_K_XL, a 9.2 GiB model, generating 20.3 tokens per second with a 32k context and q4_0 cache; a 7,775-token answer ran at 19.7 tokens per second. Qwen3.8-27B UD-IQ2_XXS, 6.8 GiB, is listed at 20.7 tokens per second with 64k context and q8_0 cache. Gemma 4 E4B QAT q4_0 is listed at 61 tokens per second, and Llama 3.2 3B Instruct Q4_K_M at 124 tokens per second. By comparison, the 27B model running on CPU as a payload and streamed from disk managed about one token per minute.
The app presents a dashboard for generation speed, resource use, temperatures, power, memory allocation and storage. Its library can load GGUF files from internal storage, M.2 or USB drives, and its model browser lists Hugging Face GGUF repositories and checks whether files fit. Downloads can resume after interruption and are checked against SHA-256. A planner selects a context length and KV-cache type based on available memory; switching models restarts the app because unloading in place leaves memory held by the pool.
A browser chat is served at port 8081, alongside OpenAI Chat Completions and Responses and Anthropic Messages APIs. The README says tools including opencode, Codex, Claude Code and aider can connect to it. Installation requires a jailbroken PS5 and components including kstuff-lite, etaHEN and ShadowMountPlus; elfldr is strongly recommended for sensors and storage data, while ftpsrv is recommended for installation and file transfer. The repository offers versioned releases and checksums, but states no price; the project is open source under GPL-3.0-or-later.
What the archive adds
New Runtime’s earlier coverage of PS5LM on 2026-10-20 described porting a general inference runtime to a constrained local-inference target. PS5LM Plus is a new development on that known story: the repository reports a native GPU application, richer model management and network APIs, rather than only a general llama.cpp port.
Trend signal
This does not yet fit a pattern New Runtime tracks.
Why now
The repository positions the app as a distinct step beyond CPU-only payload experiments: moving llama.cpp into a native title unlocks the GPU and more memory. Its reported comparison—about one token per minute for the 27B CPU payload versus roughly 20 tokens per second in the native app—makes that implementation change legible.
Why it matters
For teams working on edge inference, PS5LM Plus demonstrates the difference between making a runtime technically execute and making it useful as an application. The console’s GPU, unified memory pool, model-fit checks, resumable downloads and API surface are packaged behind one interface. That is a useful engineering pattern: expose hardware constraints and operational state rather than leaving users to guess whether a model will load.
The trade-off is significant. This depends on a jailbroken console, firmware-sensitive homebrew components and a storage arrangement outside ordinary consumer software. Its speed table is evidence for this specific setup, not a general claim about consoles or local inference. Builders can still study its architecture for other constrained devices: port a supported general runtime, measure representative models, and expose memory, context and cache limits to users.
What happens next
- PS5LM Plus will publish additional versioned releases or compatibility guidance as its native app matures (6 months, 60%). Watch for: new tagged releases or README notes documenting supported firmware, models or installation changes.
Limits and uncertainty
The performance figures are reported by the project and are not independently benchmarked here; the README does not describe a standardized comparison methodology or repeated-run variance. Results depend on model quantization, context, KV-cache choice, storage and console setup. The repository describes compatibility and installation requirements, but performance on other firmware versions or hardware revisions is not established by the listed measurements. A jailbroken PS5 and its required payloads also limit who can reproduce the results. The project states that it is not affiliated with Sony or Alibaba.
Skill to build now
If you are evaluating local inference on constrained hardware, build a small model-fit and measurement table before optimizing. Record the exact checkpoint and quantization, context length, cache type, storage source and generation speed, then expose those constraints in the interface. PS5LM Plus’s planner and per-model fit checks show how that information can become part of the user experience instead of a setup note.
Reader takeaway
PS5LM Plus turns llama.cpp on a jailbroken PS5 into a GPU-backed local app with a dashboard and network API. Its speed figures are striking compared with the CPU payload, but the practical lesson is the packaging of runtime, hardware limits and model selection—not that a standard PS5 is ready to become an inference server.