codecraftersllc/ ornith-1.5-35b-a3b-abliterated-q2:latest

129 2 weeks ago 

ollama run codecraftersllc/ornith-1.5-35b-a3b-abliterated-q2

Details

2 weeks ago

6f19f8350c75 · 13GB

model
qwen35moe
·
35.5B
·
Q2_K
params
{ "num_ctx": 4096, "temperature": 0.6, "top_k": 20, "top_p": 0.95 }

Readme

Ornith-1.5-35B-A3B Abliterated — CRACK Q2_K

An independently tuned, imatrix-assisted Q2_K GGUF of Ornith-1.5-35B-A3B, published as a separate low-VRAM Ollama repository.

This repository is distinct from the existing codecraftersllc/ornith-1.5-35b-a3b-abliterated listing. The existing Q3 listing is preserved; this repository contains the DealignAI CRACK Q2_K artifact.

Quick start

ollama run codecraftersllc/ornith-1.5-35b-a3b-abliterated-q2

The explicit quantization tag is also available:

ollama run codecraftersllc/ornith-1.5-35b-a3b-abliterated-q2:q2_k

To disable thinking for a direct response:

ollama run codecraftersllc/ornith-1.5-35b-a3b-abliterated-q2 --think=false

Model details

  • Architecture: qwen35moe
  • Total parameters: 35.5B
  • Quantization: Q2_K
  • Native context length: 262,144 tokens
  • Ollama default context: 4,096 tokens
  • Ollama capabilities: text completion, thinking, and tools
  • Ollama input: text only
  • Published model size: approximately 13 GB

The default context is intentionally conservative because KV-cache memory grows with context length. Increase it only after validating memory use on your hardware.

Hardware notes

The local validation run used Ollama 0.33.2 on an Apple M3 Max. At num_ctx=4096, ollama ps reported the model at 100% GPU residency and 13 GB loaded size. A dedicated 16 GB VRAM card should be tested independently; GPU memory behavior varies by backend, context length, runtime version, and concurrent workloads.

Vision limitation

The upstream release includes a separate F16 vision projector, mmproj-Ornith-1.5-35B-A3B-f16.gguf. Ollama’s documented GGUF import path does not attach that file as a normal adapter, and the tested package does not advertise vision. This repository is therefore published as text-only. Do not assume image support from the underlying upstream model card.

Quality tradeoff

Q2_K substantially reduces memory use but also reduces quality compared with higher-precision quants. The upstream DealignAI card reports MMLU of 69.8% for its CRACK Q2_K build versus 77.9% for the base model at the same quantization. Those figures are upstream-reported and were not independently reproduced here.

Provenance

The language model file was downloaded from:

The local language-file SHA-256 verified before packaging was:

b171351fe1feda16d0a33cc71da256aac0dd711af7479c075ca4bea41c596526

Please credit the upstream DealignAI model card, the Ornith model authors, and any contributors identified there. Review the upstream license and usage terms before redistribution or commercial use.

Responsibility

This is an abliterated/uncensored model variant and may produce unsafe, inaccurate, or otherwise inappropriate output. It does not provide refusal or safety guarantees. Users are responsible for prompts, outputs, deployment, and compliance with applicable policies and laws.