codecraftersllc/ ornith-1.5-35b-a3b-abliterated:q3_k_m

768 1 week ago

vision tools thinking 
ollama run codecraftersllc/ornith-1.5-35b-a3b-abliterated:q3_k_m

Details

1 week ago

26c4fcbf76dc · 18GB ·

model
qwen35moe
·
34.7B
·
Q3_K_M
projector
clip
·
447M
·
BF16
license
Apache License 2.0 Derived from ornith-ai/Ornith-1.5-35B-A3B. Abliterated + Dynamic GGUF by alztrk:
params
{ "num_ctx": 32768, "stop": [ "<|im_end|>", "<|im_start|>" ], "tempe

Readme

Ornith-1.5-35B-A3B-Abliterated (Q3_K_M)

Ollama packaging of the Dynamic Q3_K_M GGUF from alztrk/Ornith-1.5-35B-A3B-Abliterated.

Please credit and see the original model card there. GGUF files live in alztrk/Ornith-1.5-35B-A3B-Abliterated-GGUF. The base model is ornith-ai/Ornith-1.5-35B-A3B.

This listing adds the official Ornith vision projector (mmproj-Ornith-1.5-35B-BF16.gguf) so Ollama exposes vision as well as thinking and tools.

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

Overview

Ornith-1.5-35B-A3B-Abliterated is a refusal-ablated variant and GGUF quantization suite of ornith-ai/Ornith-1.5-35B-A3B.

It is created through directional ablation surgery on hidden state activations across all 40 layers, orthogonally projecting refusal directions out from attention and MLP down-projection matrices.

The Hugging Face repository provides both the full-precision Safetensors checkpoint and the Dynamic GGUF quantization suite.

Ornith-1.5-35B-A3B is a reasoning model: by default the assistant turn opens with a <think> … </think> block before the final answer. It is also a vision-language MoE (text + image).

Model architecture

  • Total parameters: 35.8B
  • Active parameters per token: ~3.1B (8 active experts + shared experts)
  • Total experts: 256
  • Layers: 40 hybrid layers (DeltaNet linear attention + full attention)
  • Context length: native 262,144 tokens (262K)
  • This Ollama tag: Dynamic Q3_K_M (~16 GB) + CLIP vision projector (~0.9 GB)

GGUF quantizations use Dynamic Hybrid K-Quantization: core attention matrices (attn.wv, attn.wo) retain higher precision while MoE routing FFNs are quantized.

File (on Hugging Face) Format Size Description
Ornith-1.5-35B-Abliterated-Dynamic-Q3_K_M.gguf Q3_K_M (Dynamic) ~15.61 GB Lower memory footprint. This listing.
Ornith-1.5-35B-Abliterated-Dynamic-Q4_K_M.gguf Q4_K_M (Dynamic) ~19.71 GB Balanced performance for 12GB–16GB VRAM GPUs.
Ornith-1.5-35B-Abliterated-Dynamic-Q5_K_M.gguf Q5_K_M (Dynamic) ~23.03 GB Higher precision retention.
Ornith-1.5-35B-Abliterated-Q8_0.gguf Q8_0 ~34.37 GB High precision reference quantization.

Recommended sampling

From the original Ornith / alztrk cards:

  • General chat: temperature=0.6, top_p=0.95, top_k=20
  • To match reported benchmarks: temperature=1.0

This Ollama model defaults to num_ctx=32768. Override at runtime if you need more of the native 262K window (KV cache grows quickly).

Disable thinking when you want a direct answer:

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

License and attribution

The user assumes full responsibility for any generated output.

If you use Ornith in research, please cite:

@misc{ornith_1_5,
    title = {{Ornith-1.5}: From Self-Scaffolding to Self-Improvement},
    url = {https://ornith.ai/ornith_1_5.html},
    author = {{Ornith Team}},
    year = {2026}
}