Ornith-1.0-35B-ThinkingCap-GGUF

Ornith-1.0-35B — Thinking-Cap — GGUF

This card describes the GGUF builds of khudgins/Ornith-1.0-35B-ThinkingCap — the ThinkingCap reasoning-efficiency fine-tune of deepreinforce-ai/Ornith-1.0-35B (a 256-expert Mixture-of-Experts coding model) — 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

Lower quants aren't provided here — on the 9B a Q4 of the base couldn't complete the eval, so Q8 is the floor I'd trust for this family.

Usage

Ollama (pulls the GGUF straight from this repo):

ollama run hf.co/khudgins/Ornith-1.0-35B-ThinkingCap-GGUF:Q8_0 "How many positive integers under 1000 are divisible by neither 5 nor 7?"

llama.cpp:

# pull + run
llama-cli   -hf khudgins/Ornith-1.0-35B-ThinkingCap-GGUF:Q8_0 -p "your prompt"
# or serve an OpenAI-compatible endpoint
llama-server -hf khudgins/Ornith-1.0-35B-ThinkingCap-GGUF:Q8_0

Fidelity

The tune's accuracy/token numbers (GSM8K 96% @ −24% tokens, HumanEval held, MBPP +2, avg 80.7% → 83.3%) were measured on the adapter via transformers — see the adapter card.

For the 9B, a matched f16-vs-Q8 Ollama sweep showed quantization was effectively free (−1.4 pts avg). For this 35B build I ran a coherence check on the Q8 rather than a full N=50 quant sweep — it reasons correctly (clean inclusion-exclusion derivations, correct answers) — so treat Q8 as the practical build and f16 as the precision reference, with the caveat that the quant delta here is un-swept rather than measured.

Notes

• No speculative decoding — the Qwen3.5 MTP head is dropped for llama.cpp compatibility.
• Reconverting from the merged HF model? On this MoE the MTP layer count lives in the nested config.text_config — set text_config.mtp_num_hidden_layers = 0 before converting, or the converter emits a phantom block (blk.40, block_count 41) and the GGUF fails to load with missing tensor 'blk.40.attn_norm.weight'. Full recipe in the tool repo (tools/).

License & provenance

MIT, same chain as the adapter: base deepreinforce-ai/Ornith-1.0-35B (MIT) → post-trained on Qwen3.5 (Apache-2.0). Method (an independent reproduction of BottleCap AI's ThinkingCap), training, and full results: see the adapter model card and the tool repo.