TokenStation5 01.004.000

What's new since 01.003.000

Google's Gemma 3 12B

• Gemma 3 runs on the PS5 (#94). Gemma 3 12B (bartowski's google_gemma-3-12b-it-Q4_K_M.gguf, 6.8 GB) passed every GPU check against the CPU and 59 of 60 app-style checks, tool calls and a write-a-file agent task included. Find it in the Models page's Hugging Face search.
• Most of its layers only look back 1,024 tokens, so they keep a small rolling memory, and it fits about 29,700 tokens of context next to its weights.
• It reads prompts at 250 tokens a second and writes about 40 a second, up from 188 and 35 in its first build on the console.

Faster and steadier

• New GPU kernels for wide attention heads and sliding windows: prompts read a third faster on Gemma, and writing is 13% faster (#94).
• Switching models no longer fails with "out of GPU memory" after a bigger model was loaded: the model's memory is now split into smaller pieces (#94).
• A GPU check on the model you're serving no longer runs out of memory (#94).

Better answers out of the box

• Known-good settings (#92): each checked model (Qwen 3 8B, Qwen 2.5 Coder 7B, Llama 3.1 8B, Llama 3.2 3B, Mistral Nemo 12B and now Gemma 3 12B) gets the sampling settings that scored best on the PS5. An app's own settings still win.
• Four fixes to how tool calls and thinking are read back from the models (#92).
• Mistral Nemo 12B takes tools in its own format (#91).
• The Models page and Hugging Face search only offer models the engine can run (#90).

Your models survive updates

• The first start after a store update brings your downloaded models back, through the jailbreak's FTP server (#93). Start the title once after updating. Settings, keys and chats were already kept.

Install

Install or update from the Prospero Store, or as before: PPSA99555.zip holds the PPSA99555/ app folder.