cff-version: 1.2.0
message: "If you use Piko-9b, please cite it as below, and cite the upstream models it is composed from."
title: "Piko-9b: a composed 9.65B hybrid-attention vision-language model"
abstract: >-
  Piko-9b is a 9.65-billion-parameter vision-language model assembled by splicing a
  fine-tuned language backbone, derived from deepreinforce-ai/Ornith-1.0-9B through a
  chain of merged QLoRA stages, with the vision tower of Qwen/Qwen3.5-9B copied
  verbatim. It uses a hybrid attention stack of 24 gated linear-attention layers and
  8 full-attention layers, with a declared 262,144-token context.
authors:
  - name: "Dexy"
type: software
license: Apache-2.0
version: "1.0.0"
date-released: "2026-07-29"
url: "https://huggingface.co/Dexy2/Piko-9b"
repository-code: "https://github.com/itsdexy/Piko-9b"
keywords:
  - vision-language-model
  - multimodal
  - hybrid-attention
  - linear-attention
  - qwen3.5
  - ocr
  - document-understanding
references:
  - type: software
    title: "Ornith-1.0-9B"
    authors:
      - name: "DeepReinforce AI"
    url: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B"
    license: MIT
    notes: "Upstream of the language backbone (9,197,093,888 parameters)."
  - type: software
    title: "Qwen3.5-9B"
    authors:
      - name: "Qwen Team, Alibaba Cloud"
    url: "https://huggingface.co/Qwen/Qwen3.5-9B"
    license: Apache-2.0
    notes: "Source of the vision tower and merger (456,010,480 parameters), copied verbatim."