license: mit
base_model: ornith-ai/Ornith-1.5-397B
base_model_relation: quantized
library_name: transformers
pipeline_tag: text-generation
language: [en]
tags: [uncensored, abliterated, nvfp4, modelopt, sglang, blackwell, qwen3_5_moe]
Ornith-1.5-397B uncensored (NVFP4)
TL;DR: Ornith-1.5-397B with refusal behavior ablated from the weights (1/1120
hard refusals cross-dataset), in NVIDIA NVFP4 experts-only format: 232 GB, serves on
4x RTX PRO 6000 Blackwell.
Same two-pass ablated master as the FP8 sibling repo, quantized with NVIDIA TensorRT
Model Optimizer 0.46 (NVFP4, experts only: attention, shared expert, router,
embeddings, linear attention and vision tower stay BF16). The practical way to run an
uncensored 397B-class model on a single workstation node.
Facts & figures
| Base model | ornith-ai/Ornith-1.5-397B |
| Method | Two-pass biprojection abliteration on BF16, then ModelOpt NVFP4 experts-only (group size 16, fp8 block scales, dynamic activation scales; 128 calibration samples, mlabonne/harmless_alpaca) |
| Size | 232 GB (from 794 GB BF16, -71%) |
| Hard refusals | 1/100 mlabonne (extended decline-phrase scorer, served NVFP4 artifact); FP8 sibling: 1/1120 cross-dataset |
| Measured throughput | 104.6 tok/s decode measured on the v1 NVFP4 build (SGLang, TP=4, single request, fp8 KV cache); not re-measured on v2, same weight shapes and kernels |
| Context length | 262,144 tokens |
| Quant + serve hardware | NVIDIA RTX PRO 6000 Blackwell 96 GB (SM120); quant on 7x, validated serving on 4x |
| Toolchain | ModelOpt 0.46.0, SGLang (sglang-omega build), transformers 5.12.1 |
Refusal evaluation
Served NVFP4 artifact, 100 mlabonne harmful prompts, temperature 0, extended
hard-decline scorer (classic + hedged decline phrases):
| Metric | This model |
|---|---|
| HARD refusals (effective rate) | 1/100 |
| position-aware markers (first 40 words) | 13/100 |
| naive markers (anywhere) | 80/100 (mostly false positives; see the FP8 repo for the full explanation and the 5-dataset breakdown) |
Benchmark results (measured on the FP8 sibling, same ablated master)
Custom zero-shot protocol, temp 0, SGLang TP2 x PP3. The FP8 and NVFP4 repos share the
identical two-pass ablated BF16 master; only the weight format differs. These numbers
were not re-run on this NVFP4 artifact.
| Suite | baseline | ablated (FP8) |
|---|---|---|
| MMLU-Pro (500 questions, test split, seed 42) | 0.810 | 0.814 |
| GSM8K (200 problems) | 0.950 | 0.950 |
| GPQA-Diamond (198 questions) | not run under this protocol | 0.631 |
| Needle passcode recall (32k / 64k / 128k) | 9/9 |
Paired per-question analysis on MMLU-Pro/GSM8K shows no measurable difference to the
baseline (MMLU delta +0.4 pp, 95% bootstrap CI [-1.8, +2.6] pp; GSM8K delta 0).
Full protocol details and the 5-dataset refusal breakdown are on the FP8 repo card.
Run it with SGLang
Validated command (TP=4 on 4x 96 GB GPUs, 56.8 GB weights per GPU):
python -m sglang.launch_server \
--model-path ressl/Ornith-1.5-397B-uncensored-NVFP4 \
--served-model-name Ornith-1.5-397B-uncensored-NVFP4 \
--host 0.0.0.0 --port 30000 \
--tp-size 4 \
--quantization modelopt_fp4 \
--trust-remote-code --dtype auto \
--context-length 262144 \
--kv-cache-dtype fp8_e4m3 \
--max-running-requests 8 \
--chunked-prefill-size 16384 \
--mem-fraction-static 0.88 \
--tool-call-parser qwen3_coder \
--reasoning-parser qwen3
Companion formats
- FP8 (405 GB, drop-in replacement for the official FP8 repo): ressl/Ornith-1.5-397B-uncensored-FP8
Quality & limitations
- Will comply with requests a stock model refuses; use responsibly and within your
local laws. Intended for security research, red-teaming and creative work. - NVFP4 is a 4-bit weight format; expect a small quality delta vs the FP8 sibling on
hard reasoning tasks. The refusal eval above was run on this exact NVFP4 artifact;
benchmark suites were run on the FP8 sibling (see its card). - Identity answers reflect the base model's training (inherited from ornith-ai), not
an artifact of abliteration or quantization. - Vision tower is included and kept in BF16; only the language decoder's routed
experts are NVFP4.
Provenance & reproducibility
Two-pass biprojection-abliterated BF16 master (details in the FP8 repo card),
quantized with mtq.NVFP4_EXPERTS_ONLY_CFG, calibration 128 samples x 512 tokens.
Export format verified structurally (packed uint8 weights, fp8 e4m3 block scales
[out, in/16], fp32 tensor scales) and validated serving with SGLang modelopt_fp4
before upload.
License & credits
MIT license, inherited from the base model by the Ornith AI authors. Abliteration,
quantization and validation by Robert Ressl
(Hugging Face · Website · LinkedIn · Patreon).
Built with NVIDIA TensorRT Model Optimizer and SGLang.
Support this work: if these models are useful to you, consider supporting on
Patreon; more at ressl.ch.
Version note (2026-08-23): v2 artifact (two-pass abliteration). Includes all
tokenizer/processor files; an earlier v2 upload missing them was replaced.