This card describes the GGUF builds of khudgins/Ornith-1.0-9B-ThinkingCap — the ThinkingCap reasoning-efficiency fine-tune of deepreinforce-ai/Ornith-1.0-9B — with the LoRA merged into the base and converted for llama.cpp / Ollama. Text-only (the base's vision tower is dropped in conversion).

Files

File	Quant	Size	Use
`Ornith-1.0-9B-ThinkingCap-Q8_0.gguf`	Q8_0	9.5 GB	Recommended daily driver — ~1.4 pts avg below f16, half the memory
`Ornith-1.0-9B-ThinkingCap-f16.gguf`	F16	17.9 GB	Publish-precision reference

Quantization is effectively free here (see the fidelity table below). Lower quants are not recommended — a Q4 of the base failed to complete the eval at all.

Usage

Ollama (pulls the GGUF straight from this repo):

ollama run hf.co/khudgins/Ornith-1.0-9B-ThinkingCap-GGUF:Q8_0 "A store had 48 apples, sold 3/4, then got 30 more. How many now?"

llama.cpp:

bash

# pull + run
llama-cli   -hf khudgins/Ornith-1.0-9B-ThinkingCap-GGUF:Q8_0 -p "your prompt"
# or serve an OpenAI-compatible endpoint
llama-server -hf khudgins/Ornith-1.0-9B-ThinkingCap-GGUF:Q8_0

Quantization fidelity

Benchmarked via Ollama (one harness for f16 and Q8, isolating the quant effect):

Format	Size	GSM8K	HumanEval	MBPP	avg
f16	17.9 GB	91.3%	90.9%	77.0%	86.4%
Q8_0	9.5 GB	90.0%	89.0%	76.0%	85.0%

Q8 lands within 1–3 problems of f16 on every benchmark with identical token behavior.

Notes

No speculative decoding — the Qwen3.5 MTP head is dropped for llama.cpp compatibility.
Reconverting from the merged HF model? Set mtp_num_hidden_layers=0 in config.json first, or the converter counts a phantom 33rd block (blk.32) and the GGUF fails to load. (Full write-up in the tool repo.)

License & provenance

MIT, same chain as the adapter: base deepreinforce-ai/Ornith-1.0-9B (MIT) → post-trained on Qwen3.5 (Apache-2.0). Method, training, and full results: see the adapter model card.