Qwen3.5-4B-pouw

A pouw shaping repo for omnipearl/Qwen3.5-4B. It contains no weights — only the metadata that makes omnipearl/Qwen3.5-4B mine OmniPearl Proof-of-Useful-Work while it serves. The base weights are pulled from the base repo on load.

OmniPearl's mining is output-preserving: generation is bit-identical to the base model. The eligible transformer matmuls (in_features == common_dim = 2560) are reused as PoW lottery tickets — you serve real text and mine on the same compute, no second matmul.

It is GPU-agnostic (portable Triton/PyTorch kernels, no CUDA build): RTX 3090 (sm86) → 5090 → H100 → B200, same code.

Mining shape

field	value
base model	omnipearl/Qwen3.5-4B
modality	text
common_dim	2560
rank	32
mine_layers	16 (overhead dial; layer count)
pipeline	vllm

Mining regime (LLM)

Text LLMs mine during prefill — when many tokens are processed at once (rows = tokens is large). Single-token decode does not mine (rows ≈ 1), so interactive chat mines far less than long-prompt or batched-prefill serving. Diffusion models mine on every forward (large token count always), so for continuous mining a diffusion model (see omnipearl/Z-Image-Turbo-pouw) is the stronger substrate; this LLM repo is for prefill-heavy / batch workloads.

Use

# Serve via vLLM with quantization="pouw" (vLLM-Omni plugin auto-registers it).
from vllm import LLM
llm = LLM(model="omnipearl/Qwen3.5-4B", quantization="pouw")   # mines on eligible matmuls while it serves
print(llm.generate("The history of money is"))    # generation is bit-identical to the base model

Notes

• The live PoW job + difficulty target always come from the chain at runtime — never baked into this repo. GPU kernels compile per-arch on first run (one-time, cached on disk).
• Published under the omnipearl organization. Base weights are the apache-2.0 mirror at omnipearl/Qwen3.5-4B; original model attribution is preserved there.

Generated by OmniPearl publish_pouw_models.py. License: MIT.