HERETIC 1.4.0MTPAPEXMIT

Ornith-1.0-35B-Heretic-MTP-APEX

Heretic Trial 63 BF16 merge · compatible MTP head · APEX GGUFs + mmproj

🐦 About this release

Ornith-1.0-35B is an agentic coding model from DeepReinforce AI, based on Qwen3.5 MoE. This experimental derivative merges the selected Heretic 1.4.0 Trial 63 LoRA into BF16 weights.

It also includes a compatible Qwen3.5 MTP head, a BF16 vision projector, and APEX mixed-precision GGUFs. Refusal-modified models can behave differently from the original; evaluate before deployment. License: MIT.

🧠 Model Details

Architecture	Qwen3.5 MoE, multimodal
Parameters	35B total, 3B active per token · 256 experts, 8 active
Layers	40 transformer layers + 1 MTP layer (785 BF16 tensors)
Context	262,144 tokens
Weights	BF16 safetensors, 17 shards; audited LoRA merge
License	MIT

⚡ Heretic Trial 63 of 120

Selected parameters: direction_index=21.66; attention o_proj range 1.55–2.05; MLP down_proj range 1.64–1.84.

Refusal prescreen / full evaluation	2 / 30 · 11 / 100
Validation KL mean / p95 / max	0.0242 / 0.1477 / 0.1648
Generation health	Passed · top-1 disagreement 2.0%

📊 BenchLocal Results (APEX-I-Compact)

Mode	ToolCall	BugFind	HermesAgent	Max / Effective
Thinking ↗	100	98	89	95.0 / 90.0
No Thinking ↗	97	97	85	92.2 / 76.2

Latest result cards: Thinking has 5 retry-penalty points; No Thinking has 16.

🚀 Usage

llama.cpp (vision + text + MTP acceleration)

This model preserves the native MTP head. To enable MTP speculative decoding, add:

--spec-type draft-mtp --spec-draft-n-max 2

Example:

./llama-server \
-m ./models/Ornith-1.0-35B-Heretic-MTP-APEX-I-Compact.gguf \
--mmproj ./models/mmproj-Ornith-1.0-35B-Heretic-MTP-BF16.gguf \
-ngl 99 \
-c 131072 \
--spec-type draft-mtp \
--spec-draft-n-max 2

The optimal --spec-draft-n-max value depends on hardware and MTP acceptance rate. Without these parameters, the model will work normally but will not use MTP acceleration.

🎛️ Recommended Settings

General / Coding

temperature=0.6, top_p=0.95, top_k=20

💡 What is APEX?

These GGUFs use APEX, an MoE-aware mixed-precision quantization method. It allocates precision according to tensor role and layer position; the MTP layer remains Q8_0 and the matching vision projector remains BF16.

📦 APEX Quantization Tiers

File	Size	Best For
`*-APEX-I-Quality.gguf`	21.87 GiB	Highest quality
`*-APEX-I-Balanced.gguf`	24.33 GiB	Best all-rounder
`*-APEX-I-Compact.gguf`	16.14 GiB	Best quality/size ratio
`*-APEX-I-Mini.gguf`	13.29 GiB	Most compact

Citation

@misc{ornith-35b,
    title = {{Ornith-1.0-35B}: Agentic Coding, Open to All},
    url = {https://deep-reinforce.com/ornith_1_0.html},
    author = {{DeepReinforce Team}},
    year = {2026}
}

SC117/Ornith-1.0-35B-Heretic-MTP-APEX-GGUF

Ornith-1.0-35B-Heretic-MTP-APEX

Links

Citation