steadfastgaze/Ornith-1.0-35B-Q4_K_M-MoEspresso

Ornith 1.0 35B Q4_K_M, MoEspresso package

This repository holds a MoEspresso package, and MoEspresso is the only engine that serves it: https://github.com/steadfastgaze/MoEspresso.

The weights are GGUF K-quant tensors repacked into safetensors shards, described tensor by tensor in a content-hashed manifest.

llama.cpp, transformers, and mlx-lm will not load these files.

Install and serving instructions live in the MoEspresso repository.

Original model: https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B

Quality

check	scope	result
Reference next-token loss (lower is better)	Fixed 13,215-token code corpus	This package: 1.1965; bartowski Q4_K_M in llama.cpp: 1.1946; difference: +0.0019 (+0.16%).
Greedy token agreement	Six fixed prompts, up to 64 generated tokens each	Five prompts matched at every position (198/198 tokens); the free-form sixth first differed at token 8.
Fixed-answer questions	15 independently verified answers	Initial run: 15/15; later fixed-seed runs: 14/15 and 15/15.
Mixed tasks	Four reasoning problems, three executable coding tasks, and two exact-match questions over a 35,729-token document	9/9: reasoning 4/4, executable code 3/3, long-context exact match 2/2. All generated programs passed their hidden tests.

The answer checks used fixed seeds and the model author's recommended sampling settings.

The numerical reference above is the source Q4_K_M GGUF, not the original bf16 model.

Speed

Measured on an Apple M3 Max with 128 GB of unified memory, q8 KV cache - MoEspresso engine.

context	prefill	decode
4K	1107.6 t/s	77.84 tok/s
8K		72.57 tok/s
37K	826.4 t/s	53.86 tok/s

Longest context reached in testing: 160,965 tokens, grown in one session.

Recipe

The package is built from three inputs.

input	source
bf16 safetensors	deepreinforce-ai/Ornith-1.0-35B
K-quant recipe GGUF	bartowski Q4_K_M
imatrix	bartowski imatrix GGUF, sha256 10129781b5b4…

qwen_kquant_recipe_report.json sidecar records the full recipe, including input hashes and per-shape codec assignments. package_manifest.json records the sha256 of every file; after downloading, moespresso-verify checks the package against it in one command.

The routed experts ship byte for byte from bartowski's Q4_K_M file: gate and up projections in q4_k on all 40 layers, down projections in q4_k on 20 layers and q6_k on the other 20. Nothing in those tensors was re-quantized.

The 252 dense tensors take a different path. They are encoded fresh from the bf16 release under bartowski's imatrix, with a heavier codec mix: 94 tensors in q8_0, 96 in q6_k, 10 in q5_k, 52 in q4_k. Another 361 small tensors (norms, biases) pass through in f32.

Files

Six safetensors shards (19.91 GiB, 733 tensors) plus the sidecars the engine reads: package_manifest.json (per-file sha256 and per-tensor decisions), package_plan.json (the allocation the writer consumed), agentic_profile.json (validated tool-calling defaults), and expert_hotlist.json (cold-start expert prewarm ranking). Tokenizer files and the chat template are vendored from the original release.

Credits

This quantization is derived from bartowski's quantization work; the expert bytes in this package ship unchanged and the imatrix guiding the dense encoding as well. The model is Ornith 1.0 by deepreinforce-ai. Everything in the chain is MIT: the original release, bartowski's GGUFs, and this package.