ornith-nvfp4/CITATION.cff at main · t-timms/ornith-nvfp4 · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
cff-version: 1.2.0
title: ornith-nvfp4
abstract: >-
  A REAP expert-pruned (50%, 256->128 experts) and GPTQ-NVFP4A16-quantized
  build of ornith-ai/Ornith-1.5-35B-A3B, with the MTP head and vision tower
  stripped, served by vLLM inside 16 GB of consumer VRAM on an RTX 5070 Ti
  (SM120). Final artifact is 12.47 GiB. Published SWE-bench Verified
  (22/50 = 44.0%, official harness), HumanEval+ (84.2%) and MBPP+ (89.2%)
  figures on the same 50-instance slice as the prior kat-coder-nvfp4 release.
  Includes an architecture-detected router-renormalization fix for the
  qwen3_5_moe REAP adapter (t-timms/reap-cuda).
type: software
message: >-
  If you use this checkpoint, please also cite the upstream
  ornith-ai/Ornith-1.5-35B-A3B model from which it is derived.
authors:
  - name: t-timms
repository-code: https://github.com/t-timms/ornith-nvfp4
url: https://huggingface.co/Ttimms/Ornith-1.5-35B-A3B-REAP-50-NVFP4A16
license: MIT
keywords:
  - expert-pruning
  - reap
  - nvfp4
  - quantization
  - mixture-of-experts
  - vllm
  - agentic-coding
references:
  - type: software
    title: Ornith-1.5-35B-A3B
    authors:
      - name: ornith-ai
    url: https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B