---
license: mit
base_model:
  - deepreinforce-ai/Ornith-1.0-35B
base_model_relation: quantized
language:
  - en
library_name: gguf
tags:
  - gguf
  - llama.cpp
  - mtp
  - multi-token-prediction
  - speculative-decoding
  - qwen35moe
  - moe
pipeline_tag: text-generation
---

# Ornith-1.0-35B-MTP-GGUF

[Ornith-1.0-35B](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B) (qwen35moe, 35B-A3B,
Qwen3.5 base) is a strong agentic-coding MoE that **ships without MTP heads**. This GGUF has an
**MTP head grafted in** so it can use llama.cpp's `--spec-type draft-mtp` self-speculative
decoding for a real speedup, with no quality change to the base weights.

Quantization: **Q6_K** (body + grafted MTP head).

## Performance (M3 Max, measured, real-prompt benchmark)

Generation speed (tg128) on a real code-continuation prompt, sweeping draft depth:

| Mode | tok/s | Speedup | MTP acceptance | mean accepted len |
|---|---|---|---|---|
| AR (no MTP) | 66.6 | 1.00× | — | — |
| **draft-mtp n_max=1** | **83.8** | **1.26×** | 92.2% | 1.92 |
| draft-mtp n_max=2 | 82.8 | 1.24× | 82.5% | 2.65 |
| draft-mtp n_max=3 | 81.7 | 1.23× | 78.2% | 3.35 |
| draft-mtp n_max=4 | 75.9 | 1.14× | 68.1% | 3.72 |

**Best: `--spec-draft-n-max 1`, ~1.26×.** (Acceptance is much higher on real text than on random
tokens — benchmark with a real prompt or you'll badly underestimate MTP.)

## Usage (llama.cpp)

```bash
llama-server -m Ornith-1.0-35B-Q6_K-MTP.gguf -ngl 99 -c 32768 \
  --spec-type draft-mtp --spec-draft-n-max 1 --port 8080
```
Requires a llama.cpp build with `draft-mtp` speculative support.

## Provenance & licensing

- **Base model**: [deepreinforce-ai/Ornith-1.0-35B](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B) — MIT.
- **MTP head**: grafted from a same-architecture (`qwen35moe`, 40 blocks) sibling that ships MTP
  heads, following the cross-model graft approach published by
  [skinnyctax/Ornith-1.0-35B-Q6_K-Frankenstein-MTP-GGUF](https://huggingface.co/skinnyctax/Ornith-1.0-35B-Q6_K-Frankenstein-MTP-GGUF) (MIT).
  The 20 MTP head tensors (`blk.40.*`, incl. `nextn.*`) are appended to the base GGUF and metadata
  patched (`block_count` +1, `nextn_predict_layers=1`).

Released under MIT. No weights retrained — this is a head graft + metadata patch.