← back to catalog · registered 2026-09-23 08:57

kataguru/Qwen3.8-27B-Finnish-Titan-v2.0-Uncensored-NVFP4

kataguru 27B multimodal second-order
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/kataguru%2FQwen3.8-27B-Finnish-Titan-v2.0-Uncensored-NVFP4"
Response includes
  • classification m-uncensored
  • files 22
  • author_summary 20 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
463
Likes
0
Model age
6w ago
created 2026-09-23
Downloads over time
Now351→from0↑0%
01292573860 on Sep 23351 on Sep 25Sep
Sep 23 → Sep 25 · 3 snapshots · spans 2 days

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 3 formats · 808 downloads combined

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

Metadata

License
apache-2.0
Languages
fi en
Tags
safetensors qwen3_5 finnish suomi nvfp4 blackwell fp4 compressed-tensors vllm reasoning uncensored multimodal

Related

Total size
26.7 GB
Files
22
Quantizations
1
Registered
2026-09-23 08:57
Last updated on HF
2026-09-25 20:18

Files by quantization

Auxiliary files 22 files 26.7 GB
model-00006-of-00007.safetensors 4.66 GB a44aff39 download
model-00002-of-00007.safetensors 4.65 GB e1b00fbc download
model-00004-of-00007.safetensors 4.65 GB ec3c1ad2 download
model-00003-of-00007.safetensors 4.65 GB 25409422 download
model-00005-of-00007.safetensors 4.64 GB a2160222 download
model-00001-of-00007.safetensors 2.37 GB c098f75a download
model-mtp.safetensors 810 MB 9f1f9e40 download
model-00007-of-00007.safetensors 191 MB af8e851a download
model-mtp-restored.safetensors 100 MB fbea2586 download
tokenizer.json 19.1 MB 06b95093 download
model.safetensors.index.json 164 KB 7001e4f6 download
config.json 29.3 KB b3ff639a download
chat_template.jinja 15.3 KB 0cc5dbb6 download
quantization.log 11.7 KB 2ac4e410 download
README.md 3.90 KB ac00b551 download
.gitattributes 1.53 KB 52373fe2 download
processor_config.json 1.16 KB 33818c7f download
tokenizer_config.json 1.10 KB d1a20cc3 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 0bc3addd download

README current version from Hugging Face


language:

  • fi
  • en
    license: apache-2.0
    base_model: kataguru/Qwen3.8-27B-Finnish-Titan-Uncensored
    tags:
  • finnish
  • suomi
  • nvfp4
  • blackwell
  • fp4
  • compressed-tensors
  • vllm
  • reasoning
  • uncensored
  • multimodal
  • vision
  • mtp
    pipeline_tag: text-generation

Qwen3.8-27B-Finnish-Titan-v2.0-Uncensored-NVFP4

1. Mitä Kataguru muutti? (Muutosyhteenveto)

Kenttä Kuvaus
Base model kataguru/Qwen3.8-27B-Finnish-Titan-Uncensored (Upstream: Qwen/Qwen3.8-27B & TWIN-TURBO-Fable-Cold-Fusion-709)
Katagurun muutos Titan v2.0 -jatkokoulutus (Leonardo 20k SFT + Epistemic Anchor SFT), Kataguru Master Chat Template v1.0, Blackwell NVFP4 (W4A4/W4-FP8) -kvantisointi
Painojen koko (Levytila) 27,7 Gt (8 safetensors-tiedostoa; FP4-painot + FP8-skaalat + BF16 MTP/visio)
VRAM-tarve (vLLM) 14,6 Gt / GPU (TP=2, esim. 2x RTX 5090) tai ~29 Gt (1x 32GB GPU)
Mitä EI muutettu 27B arkkitehtuuri, perusmorfologia, huolella valitut kvantisoimattomat kerrokset (lm_head, mtp, linear_attn, visual)
Ensisijainen käyttö Huippunopea tuotantopäättely NVIDIA Blackwell -arkkitehtuurilla (RTX 50 -sarja, B200): maksimaalinen läpimeno ja minimaalinen latenssi
Suositeltu versio NVFP4 (compressed-tensors), 2x RTX 5090 / TP=2
Testattu laitteisto 2x NVIDIA GeForce RTX 5090 32GB (vLLM v0.26, TP=2, CUDA 13.2)

2. Empiirinen Vertailutaulukko: Titan 1.0 vs. Titan v2.0

Testi / Osa-alue Titan 1.0 (Syyskuu 2026) Titan v2.0 (Uusi versio) Muutos (Delta) & Huomiot
Hallusinaatioiden torjunta (20 ansaa) 85,0 % (17/20) 90,0 % (18/20) +5,0 % (Torjui Mannerheim-Heisenbergin, Päätalo-algoritmin, sunnuntaipesulait, keksityt API:t)
FinnishBench-Kataguru-50 (AWQ / NVFP4) 163,8 tok/s (AWQ) / 137,8 tok/s (NVFP4) 140,1 tok/s (AWQ) / 114,0 tok/s (NVFP4) Molemmat mallit saavuttavat huippuluokan läpimenon (MTP $K=2$ + FlashInfer XQA). Ero benchmarkissa selittyy vastauspituudella (Titan 1.0 keskim. 546/572 tok vs. v2.0 tiivis 366/297 tok). Natiivi BF16-nopeus oli Titan 1.0:lla 44,8 tok/s ja Titan v2.0:lla 50,7 tok/s (+13,2 %).
Koodaustesti (LFU Cache $O(1)$, C11, SQL) 8,8 / 10 9,8 / 10 Täysi $O(1)$ LFU Python-toteutus dummy-päillä, BST-bugikorjaus, optimaalinen window SQL
Vaativa suomi (Volter Kilpi, anglismien karsinta) Hyvä Erinomainen 137 sanan katkeamaton saaristolaisvirke, piiloanglismien eliminointi
Päättelyarkkitehtuuri & Chat Template Erillinen tturbo Kataguru Master v1.0 Yhdistetty {REASON:spoon/einstein/xhigh/low/none} oletuksena
Sensuroimattomuus & Aikuiskerronta Täysin avoin Täysin avoin Fiktiivinen konteksti poistaa kieltäytymisen; ei moraalisaarnoja
Koulutuksen lopullinen loss ~0,68 0,5879 Leonardo 20k + Epistemic ankkurointi

3. NVFP4 (NVIDIA Blackwell Native FP4) -arkkitehtuuri

  • Natiivi 4-bittinen liukulukulaskenta: NVFP4 hyödyntää Blackwell-arkkitehtuurin (kuten RTX 50 -sarjan ja B200) uusia Tensor-ytimiä ja FP4-skaalauksia (scale_dtype: torch.float8_e4m3fn).
  • Muistijalanjälki: Mallin painot mahtuvat alle 15 gigatavuun VRAM-muistia.
  • Spekulatiivinen MTP-dekoodaus: MTP-pää säilytetty ja aktivoitu $K=2$ spekulatiiviseen generointiin vLLM:ssä.

4. Tuotantoasennus (vLLM)

vllm serve kataguru/Qwen3.8-27B-Finnish-Titan-v2.0-Uncensored-NVFP4 \
  --tensor-parallel-size 2 \
  --quantization compressed-tensors \
  --kv-cache-dtype fp8 \
  --enable-chunked-prefill \
  --speculative-config '{"method":"mtp","num_speculative_tokens":2}' \
  --limit-mm-per-prompt '{"image":99}' \
  --max-model-len 32768 \
  --max-num-batched-tokens 8192 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder \
  --override-generation-config '{"temperature":0.6,"top_p":0.95}' \
  --host 0.0.0.0 \
  --port 8000

README history 8 versions

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-09-24Update Model Tree lineage and YAML base_model metadata249378e3.7 KB
    Loading...
  2. 2026-09-23Fix Titan 1.0 baseline speed comparison table (apples-to-apples)9b55f633.8 KB
    Loading...
  3. 2026-09-23Update model card with Titan v2.0 banner artwork10e02b33.6 KB
    Loading...
  4. 2026-09-23docs: add explicit model weight sizes in GB and VRAM requirements to summary ...d104b103.9 KB
    Loading...
  5. 2026-09-23docs: add measured NVFP4 benchmark results47e9f253.7 KB
    Loading...
  6. 2026-09-23docs: update empirical speed benchmark with exact measured tok/sb6b09b73.7 KB
    Loading...
  7. 2026-09-23Fix speed comparison: clarify apples-to-apples baseline7269d233.6 KB
    Loading...
  8. 2026-09-23Release Qwen3.8-27B-Finnish-Titan-v2.0-Uncensored NVFP4 (Blackwell native 4-b...b2a89d93.5 KB
    Loading...
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 Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration