ProsperoAI-Qwen3.5-9B-Q4_0-PS5

ProsperoAI Qwen3.5 9B Q4_0 for PS5

This repository contains a curated, install-ready conversion of Qwen3.5-9B Q4_0 for the ProsperoAI native PlayStation 5 GPU runtime. It provides offline text chat in ProsperoAI and does not use a cloud API.

This is a runtime-format conversion, not a fine-tune. The converted weights come from unsloth/Qwen3.5-9B-GGUF, which is based on Qwen/Qwen3.5-9B. The source model and conversion are distributed under the Apache 2.0 license.

Target runtime

• Application: ProsperoAI (PPSA99004)
• Processor: PlayStation 5 AMD GPU through the native AGC compute path
• Runtime profile: prosperoai-qwen35-9b-v1
• Container format: P5LM v1, runtime layout qwen35-9b-runtime-v1
• Quantization: the source GGUF Q4_0 profile, including its mixed Q4_1, Q8_0, Q5_K, and Q6_K tensors
• Execution: GPU-resident weights and recurrent/KV state; CPU orchestration and tokenization
• Current app context limit: 4,096 tokens
• Current decoding: greedy token selection

This P5LM file is not a GGUF file and is not directly compatible with llama.cpp, ROCm, CUDA, or desktop model runners. It is laid out specifically for ProsperoAI's native PS5 AGC kernels.

Validated architecture

The converter and runtime validate this exact profile before accepting a model:

• 32 hybrid layers: 24 recurrent layers and 8 full-attention layers
• embedding width 4,096 and feed-forward width 12,288
• 16 attention heads, 4 KV heads, and head dimension 256
• vocabulary size 248,320
• RoPE sections 11/11/10 and frequency base 10,000,000
• recurrent state size 128, 16 groups, and convolution width 4
• 427 tensors in the expected Qwen3.5-9B layout

ProsperoAI currently uses the text portion of Qwen3.5-9B. Vision input and the separate multimodal projector are not included in this package.

Install

Download the repository, then copy the complete folder qwen3.5-9b-q4-0 into the app's models directory:

PPSA99004/
└── models/
    └── qwen3.5-9b-q4-0/
        ├── model.ps5lm
        ├── tokenizer.ps5tok
        └── model.json

Keep the generic filenames unchanged. The named parent folder identifies the model and allows ProsperoAI to discover multiple installed models. Restart the app after copying or replacing a model. Use Workshop to switch between this model and another supported architecture such as Mistral 7B.

Files and integrity

Source GGUF:

• Repository: unsloth/Qwen3.5-9B-GGUF
• File: Qwen3.5-9B-Q4_0.gguf
• Revision: 3885219b6810b007914f3a7950a8d1b469d598a5
• SHA-256: 17670346b4260ddcb0173965145155885024f3c9a4a24389a3370751edbcde24

The full machine-readable conversion record is retained separately at provenance/model-preparation.json; it is not required on the PS5.

Reproduce the conversion

Download the exact source GGUF listed above and the ProsperoAI Alpha 2 model tools, then run:

prepare-model --provenance provenance/model-preparation.json \
  Qwen3.5-9B-Q4_0.gguf

The recipe contains the complete validated tensor layout. The tool verifies the source and all three generated files against their recorded SHA-256 values.

PS5 validation

The native PS5 GPU runtime generated the coherent response:

Hello! I'm ProsperoAI, running locally on your PlayStation 5. How can I help you today?

It also passed an in-process Qwen → Mistral → Qwen switch. After switching back, Qwen reproduced the expected first token 9419 for the canonical chat prompt. The runtime releases the inactive architecture's large direct-memory mapping during a switch.

The current Qwen backend cold-loads in about 4 seconds and decodes at roughly 30 tokens per second at short context. A representative 44-token prompt prefilled in 1.46 seconds. Attention still becomes slower as context grows, while matching conversation prefixes are reused.

Limitations

• Requires a PS5 environment capable of running the ProsperoAI native app.
• Only the exact architecture and tensor profile above is currently supported.
• Text inference only; the vision projector and image input are unsupported.
• The app's 4,096-token limit is lower than the upstream model's limit.
• Long prompts are substantially slower than short prompts.
• Follow the upstream Qwen model card, license, and acceptable-use guidance.

Curated and converted for ProsperoAI by BlackBearReloaded.