heretic MTP Vision MIT

🐦 About Ornith

Ornith-1.0-9B is a self-improving agentic coding model from DeepReinforce AI, post-trained on top of Qwen3.5 (9B Dense) with RL to jointly optimize scaffold generation and solution rollouts.

It achieves strong performance on Terminal-Bench 2.1, SWE-Bench Verified, NL2Repo, and OpenClaw among open-source models of comparable size.

This GGUF package includes heretic ARA-LoRA abliteration for uncensored use, MTP layers injected from Qwen3.5-9B, and the mmproj-F16 vision projector for multimodal capabilities. License: MIT.

🧠 Model Details

Architecture	Qwen3.5 Dense
Parameters	~9B (all parameters active)
Layers	33 transformer layers + 1 MTP layer
Context	262,144 tokens
Attention	16 heads, 4 KV heads (GQA)
MTP	1 MTP layer, injected from Qwen3.5-9B (same architecture, compatible weights)
Thinking	Yes ( blocks)
License	MIT

🔓 heretic ARA-LoRA Abliteration

This model is abliterated using heretic with the ARA-LoRA method (Arbitrary-Rank Ablation with LoRA). ARA-LoRA identifies and removes refusal behavior by ablating specific directions in the model's weight space while preserving general capabilities through KL divergence control.

Key ablation parameters (Trial 76 of 250, best result):

Target Layers	Layers 14–16
preserve_good_behavior_weight	0.7319
steer_bad_behavior_weight	0.0001
overcorrect_relative_weight	1.1086
neighbor_count	7
Result	KL divergence: 0.0288 (< 0.05 ✅), Refusals: 3/100 (< 10 ✅)

Quantization: bnb_4bit · Batch size: 32 · Target: KL < 0.05, Refusals < 10/100

⚡ MTP (Multi-Token Prediction)

MTP layers are injected from the original Qwen3.5-9B base model (same architecture, compatible weights). MTP enables the model to predict multiple future tokens simultaneously, improving generation speed and coherence.

The MTP layer includes mtp.fc.weight and mtp.layers.0.* tensors, added on top of the 33 standard transformer layers.

Requires --chat-template chatml for proper thinking mode rendering in llama.cpp.

📊 BenchLocal Results (Q6_K, 7.6 GB)

Metric	ToolCall-15	BugFind-15	HermesAgent-20	Max	Eff.
Score	100	94	79	89.8	68.8

RTX 5070 Ti · 21 total retries · ToolCall perfect 100/100 🏆

📦 GGUF Quantizations

File	Size	Best For
*-BF16.gguf	18.4 GB	Full precision source
*-Q8_0.gguf	9.8 GB	Near lossless, highest quality
*-Q6_K.gguf	7.6 GB	Recommended balance ⭐
*-Q4_K_M.gguf	5.8 GB	Good quality, smaller size
mmproj-*-BF16.gguf	0.9 GB	Vision encoder (for image inputs)

Standard llama-quantize from BF16 source. No imatrix used — broad compatibility.

🚀 Usage

llama.cpp (text only)

./llama-server -m Ornith-1.0-9B-heretic-MTP-Q6_K.gguf --chat-template chatml -ngl 99 -c 8192 --temp 0.6 --top_p 0.95 --top_k 20

llama.cpp (vision + text)

./llama-server -m Ornith-1.0-9B-heretic-MTP-Q6_K.gguf --mmproj mmproj-Ornith-1.0-9B-heretic-MTP-BF16.gguf --chat-template chatml -ngl 99 -c 8192

🎛️ Recommended Settings

Parameter	Value
temperature	0.6
top_p	0.95
top_k	20

From official DeepReinforce AI model card.