# Benchmark Notes

Parent model: [deepreinforce-ai/Ornith-1.0-35B](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B)

Required runtime: [turbo-tan/llama.cpp-tq3](https://github.com/turbo-tan/llama.cpp-tq3)

The published comparison uses the post-fix BenchLoop results for the Ornith rows.

| Model | Field overall | Toolcall | EasyCode | Data extract | Instruct follow | Reason math | Size | Gen tok/s |
|---|---:|---:|---:|---:|---:|---:|---:|---:|
| Ornith-1.0-35B TQ3_4S | 95.75 | 88.3% | 100.0% | 86.5% | 65.5% | 73.3% | 13.00 GiB | 146.3 |
| Ornith-1.0-35B Q4_K_M | 90.17 | 86.7% | 100.0% | 86.6% | 60.0% | 73.3% | 20.00 GiB | 164.1 |

Hard86 results:

- Ornith TQ3_4S: 81.4%, 135.8 tok/s
- Ornith Q4_K_M: 82.6%, 164.4 tok/s

Field scoring formula: `0.85 * task_score + 0.15 * size_factor`, with size normalized to the smallest displayed 35B model.

BenchLoop submission was disabled. The published data-extract result includes the fence-safe JSON-wrapper parsing correction used for the Ornith comparison.