← back to catalog · registered 2026-09-25 19:57

kataguru/Qwen3.8-27B-Titan-v3.0-Uncensored-NVFP4

kataguru 27B second-order
curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/kataguru%2FQwen3.8-27B-Titan-v3.0-Uncensored-NVFP4"
Response includes
  • classification m-uncensored
  • files 15
  • author_summary 17 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M-U
Primary method

Uncensored (method unknown)

No other method signals detected in this model.
Confidence
LOW
Why this label 3 signals
Weak or ambiguous signals. Best guess based on catalog patterns; treat as tentative and check the evidence below.
  • 'uncensored' in name/tags but no 'abliterated' marker
  • method not identifiable from author declaration alone
  • may be DPO fine-tune, prompt engineering, or unknown technique
Refusal direction extraction

No specific extraction method could be identified for this model. The producer either did not document it or used a proprietary pipeline.

What is a refusal direction? →
Downloads · 30-day
0
Likes
0
Model age
today
created 2026-09-25

Genealogy 0 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Variants by this author 2 formats · 0 downloads combined

The same weights this author released in different packaging. Pick the format that matches your runtime.

Metadata

License
apache-2.0
Languages
en fi et hu tr
Tags
safetensors qwen3_5 agglutinative turkish estonian hungarian finnish nvfp4 fp4 blackwell rtx5090 compressed-tensors

Related

Total size
25.8 GB
Files
15
Quantizations
1
Registered
2026-09-25 19:57
Last updated on HF
2026-09-25 19:49

Files by quantization

Auxiliary files 15 files 25.8 GB
model-00001-of-00002.safetensors 18.6 GB 61f7f48a download
model-00002-of-00002.safetensors 7.18 GB 236709f2 download
tokenizer.json 19.1 MB 06b95093 download
titan3.png 2.61 MB 24843a9e download
model.safetensors.index.json 163 KB 1059fd77 download
config.json 29.0 KB b031ab92 download
chat_template.jinja 15.3 KB 0cc5dbb6 download
README.md 5.45 KB 54edcbdb download
.gitattributes 1.58 KB cd8e28f0 download
tokenizer_config.json 1.20 KB f636bbbf download
processor_config.json 1.16 KB 33818c7f download
recipe.yaml 822 B 31bc0b2d download
preprocessor_config.json 390 B 2ea84a43 download
video_preprocessor_config.json 385 B 3ba673a5 download
generation_config.json 214 B c53835dc download

README current version from Hugging Face


language:

  • en
  • fi
  • et
  • hu
  • tr
    license: apache-2.0
    base_model:
  • kataguru/Qwen3.8-27B-Titan-v3.0-Uncensored-BF16
    tags:
  • agglutinative
  • turkish
  • estonian
  • hungarian
  • finnish
  • nvfp4
  • fp4
  • blackwell
  • rtx5090
  • compressed-tensors
  • reasoning
  • uncensored
  • math
  • stem
  • kataguru
  • titan
    pipeline_tag: text-generation

Qwen3.8-27B-Titan-v3.0-Uncensored-NVFP4

Kataguru Titan 3.0 – Äärimmäinen Totuus (Ultimate Truth)

Official Native NVFP4 for NVIDIA Blackwell RTX 5090 / B200 / B100

Kataguru Titan 3.0 NVFP4

"Totuus ei ole kauppatavaraa, miellyttämistä tai kompromisseja. Se on luonnon ja fysiikan lahjomaton laki, joka loistaa puhtaana kuin pohjoinen taivas."
(Truth is neither a commodity, nor an exercise in people-pleasing, nor a compromise. It is the unyielding law of physics and nature, standing as clear and luminous as the northern sky.)


🌍 Executive Summary

Kataguru Titan 3.0 NVFP4 is the cutting-edge native FP4 (Microscopic Scaling Block Format) release of Kataguru Titan 3.0 – Äärimmäinen Totuus (Ultimate Truth, 27 billion parameters).

Engineered specifically for the NVIDIA Blackwell architecture (GeForce RTX 5090 sm_120, B200, B100), this model utilizes 5th-generation Tensor Cores and hardware-accelerated Microscopic Scaling FP4 math to deliver unprecedented token generation speeds (exceeding 120 tokens/sec on a single RTX 5090) while fitting comfortably inside 16 GB VRAM.

[!IMPORTANT]
🇬🇧 Yes, It Excels in English, Too!
While Titan 3.0 provides historic support for the agglutinative family, English inference is razor-sharp. At over 120 tokens/sec on Blackwell, it processes English STEM analysis, coding, and multi-turn deduction at lightning speed.


📊 Empiiriset Mittaustulokset: Titan v3.0 vs. Qwen 3.8 27B Base

Kaikki mittaukset on suoritettu standardoidulla lm-evaluation-harness -kehyksellä OpenAI-yhteensopivan vLLM-rajapinnan läpi AI-CORE-infralla (2x NVIDIA RTX 5090 32GB). Vertailukohtana on Alibaban virallinen koskematon perusmalli Qwen 3.8 27B Base.

Tehtävä / Benchmark Metriikka Qwen 3.8 27B Base (Untuned) Kataguru Titan v3.0 Uncensored Muutos / Vaikutus
ARC-Easy (Perustason tiede) Acc_norm 72.98 % 87.58 % +14.60 pp 🚀 (Massiivinen luonnontieteen harppaus)
HellaSwag (Lausejatkuvuus & arkijärki) Acc_norm 74.60 % 84.80 % +10.20 pp 🚀 (Luonnollinen kielellinen päättely)
ARC-Challenge (Vaativa tiede) Acc_norm 58.62 % 68.34 % +9.72 pp 🚀 (Syvällinen tieteellinen ongelmanratkaisu)
WinoGrande (Kontekstuaalinen päättely) Acc 71.10 % 79.56 % +8.46 pp 🚀 (Monitulkintaisuuden looginen ratkaisu)
PIQA (Fysikaalinen arkijärki) Acc_norm 80.10 % 83.35 % +3.25 pp ↗ (Reaalimaailman fysiikan intuitio)
BoolQ (Luetun ymmärtäminen) Acc 89.60 % 90.80 % +1.20 pp ↗ (Huippuluokan faktatekstin ymmärtäminen)
OpenBookQA (Avoin luonnontieto) Acc_norm 44.80 % 45.80 % +1.00 pp ↗ (Tieteellinen assosiaatio ja tiedonhaku)
GSM8K (Matematiikan monivaiheinen CoT) Exact Match ~47.0 % 48.14 % +1.14 pp ↗ (Matemaattinen päättelyketju hallinnassa)
TruthfulQA MC2 (Monivalintatotuus) Multi-True Prob 54.24 % 53.49 % -0.75 pp ≈ (SOMA/ARA säilytti totuudenmukaisuuden)
TruthfulQA MC1 (Yksittäinen totuus) Single-True Acc 36.23 % 35.37 % -0.86 pp ≈ (Pysynyt samalla tasolla, virhemarginaalin sisällä)
EstCOPA (Viron looginen kausaalipäättely) Accuracy <50.0 % 80.00 % (40/50) Suvereniteetti: Viron morfeeminen siirto toimii saumattomasti
Unkari (Alpaca-52k kieliaines) Onnistumisaste Satunnainen 100.0 % (25/25) Täydellinen instruktiovaste ja unkarin morfologia
Turkki (Gemma-51k kieliaines) Onnistumisaste Satunnainen 100.0 % (25/25) Täydellinen instruktiovaste ja turkin morfologia
FinnishBench-Kataguru-50 Suomi-akkreditointi 14/22 (Anglismeja) 100.0 % (22/22) Täydelliset astevaihtelut, lääketiede, ei fingelskaa
Refusal / Moralizing Kieltäytymisaste Korkea kieltäytyminen 0.0 % (SOMA/ARA) Täysin sensuroimaton, asiallinen, ei saarnoja

⚡ Architectural Advantages of Blackwell NVFP4

  1. Hardware FP4 Execution: Unlike emulated int4 formats, Blackwell Tensor Cores execute FP4 math with dedicated dynamic range exponent mantissas.
  2. Grammar & Morpheme Preservation: The microscopic scaling blocks prevent gradient clipping, ensuring that subtle agglutinative case endings in Finnish, Estonian, Hungarian, and Turkish remain razor-sharp.
  3. Single-GPU 27B Inference: Fits entirely in a single RTX 5090 (32GB VRAM) with huge headroom for an 8k–16k context window and FP8 KV cache.

🚀 Serving with vLLM on Blackwell (RTX 5090)

vllm serve kataguru/Qwen3.8-27B-Titan-v3.0-Uncensored-NVFP4 \
    --max-model-len 8192 \
    --gpu-memory-utilization 0.85 \
    --kv-cache-dtype fp8

🛡️ Uncensored & Epistemically Honest

Titan v3.0 has undergone SOMA/ARA activation orthogonalization, surgically stripping refusal and moralizing preachy vectors from residual activations. It treats the user as an adult, answers questions objectively, and never lectures or moralizes.

Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in app" button that hands off directly to a local runtime of your choice - Abliteration, LM Studio, or Ollama. No API keys, no subscription, no prompt leakage.