---

license: apache-2.0

language:

• fi
• en
• et
• hu
• tr

library_name: gguf

pipeline_tag: image-text-to-text

base_model:

• Qwen/Qwen3.5-35B-A3B
• kataguru/Kataguru-Sceptic-35B-A3B-v2.7-NVFP4

tags:

• gguf
• llama.cpp
• imatrix
• apex
• moe
• finnish
• coding
• swe-bench
• plc
• cnc
• wet-lab
• chemistry
• patent-ipr
• medical
• multimodal
• vision
• tool-calling
• bfcl
• uncensored
• kataguru

---

Kataguru Sceptic 35B-A3B v2.8.1 — GGUF APEX Trio (APEX-Small / APEX-Medium / APEX-Good + Vision mmproj-f16)

Kataguru-Sceptic-35B-A3B-v2.8.1-GGUF is the official imatrix-calibrated GGUF Trio + Multimodal Vision Projector (mmproj-f16) release of the Kataguru Sceptic 35B-A3B v2.8.1 Master (CP-300) Sovereign Multimodal 254-Expert Mixture-of-Experts (MoE) flagship model for llama.cpp, LM Studio, and Ollama on single-GPU (16 GB–24 GB VRAM) and workstation setups.

Built through 254-Expert Full-MoE Surgical QLoRA (r=16, alpha=32, all 40 layers, all 254 routed experts + shared experts + attention/DeltaNet projections trained simultaneously) on top of the Kataguru Sceptic 35B-A3B v2.7 Master (254-MoE) spine, v2.8.1 sets a new all-time Kataguru MoE record on the 11-Task Cognitive Suite (72.35 %), surges +4.57 pp on LiveCodeBench v6 (32.00 % Pass@1), reaches 82.33 % on BFCL V4 Single-Turn AST (+8.90 pp over v2.7), and outperforms Ornith 1.5 35B-A3B on 12 out of 15 benchmarks (+3.75 pp overall 15-benchmark mean, +4.72 pp 11-task cognitive mean, +9.02 pp BFCL V4 AST, +4.00 pp LiveCodeBench v6) while using -387M fewer parameters (34.67B vs. 35.06B).

---

0. Included GGUF Quantizations & VRAM Footprint

All three GGUF quantizations are calibrated with a bilingual (Finnish + English + STEM + Code + PLC/CNC + Wet-Lab) importance matrix (imatrix.dat) and include the native 27-layer Vision encoder projector (mmproj-Kataguru-Sceptic-35B-A3B-v2.8.1-f16.gguf):

| GGUF File | Quantization Profile | Size (GiB) | Recommended GPU VRAM | Target Use Case |

| :--- | :--- | :---: | :---: | :--- |

| Kataguru-Sceptic-35B-A3B-v2.8.1-APEX-Small.gguf | IQ2_XXS + imatrix (Q8_0 routers/attention/norms) | 11.59 GiB | 16 GB VRAM (RTX 4080 / 5080) | High-speed local agent & coding on 16 GB GPUs |

| Kataguru-Sceptic-35B-A3B-v2.8.1-APEX-Medium.gguf | IQ3_XXS + imatrix (Q8_0 routers/attention/norms) | 13.52 GiB | 16–24 GB VRAM (RTX 4090 / 5080) | Optimal quality-to-VRAM ratio for daily engineering |

| Kataguru-Sceptic-35B-A3B-v2.8.1-APEX-Good.gguf | IQ4_XS + imatrix (Q8_0 routers/attention/norms) | 15.88 GiB | 24–32 GB VRAM (RTX 3090 / 4090 / 5090) | Near-lossless reference GGUF for maximum reasoning |

| mmproj-Kataguru-Sceptic-35B-A3B-v2.8.1-f16.gguf | Native FP16 Multimodal Vision Projector (333 tensors) | 0.84 GiB | +0.9 GB VRAM (when vision enabled) | Technical schematics, P&ID diagrams, OCR, lab imagery |

---

1. Ten Sovereign Capabilities & Trained Specializations (What Sets Kataguru Sceptic v2.8.1 Apart)

Unlike standard general-purpose fine-tunes of Qwen3.5-35B-A3B (such as Ornith 1.5), Kataguru Sceptic 35B-A3B v2.8.1 combines architectural MoE surgery with eight specialized engineering, scientific, and sovereign domain curricula:

1.1 Lighter 254-MoE Architecture (-387M Parameters / -774 MB BF16 / -194 MB 4-Bit)

• Architectural Pruning of Redundant Experts: Two redundant routed experts per MoE layer (256 -> 254 experts across all 40 layers = 80 redundant experts removed) were surgically pruned and the router gates re-indexed (num_experts = 254, top_k = 8).
• Smaller & Faster: Total parameter count is reduced from 35.06B -> 34.67B (-387,082,240 parameters). Despite being smaller on disk and in VRAM, Sceptic v2.8.1 beats the larger 256-expert Ornith 1.5 by +3.75 pp across the 15-benchmark suite.

1.2 Six Native Operating Modes Built Into chat_template.jinja

Kataguru Sceptic v2.8.1 is both a flagship reasoning model and the autonomous engine of the Kataguru 4-Stage Data Factory. Its chat_template.jinja supports 6 native mode switches out of the box:

1. mode="chat" (Default — Spoon/Einstein First-Principles Mode): Direct, sycophancy-free technical execution. First-principles derivation (Einstein) + ruthless assumption stripping (Spoon).
2. mode="decision" (System 1 Zero-Output-Token Logit Firewall): Single-token (0 or 1) logprob classifier (<20 ms latency, >5M samples/day) for rapid epistemic triage ("Does this require deep System 2 reasoning or contain novel delta?").
3. mode="topic" (Deterministic 16-Category Taxonomy Classifier): Single-token (A–P) domain router across 16 industrial, scientific, coding, medical, and legal categories.
4. mode="inspect" (Stage 2 Quality Inspector Oracle): Strict JSON schema auditor scoring datasets on score (1–10) and learning_value (1–10) (>= 8.0 Golden Gate), checking math/code correctness, truncation, and refusal status.
5. mode="curate" (First-Principles Dataset Rewriter): Transforms noisy raw Q&A pairs into verified, zero-fluff -golden training samples.
6. mode="surgeon" (Autonomous 3-Stage Code & Repo Surgeon): Structured Stage 1 Fault Localization -> Stage 2 Root-Cause Patch Generation (diff) -> Stage 3 Edge-Case Verification mode for SWE-bench and Terminal-Bench repositories.

1.3 Competitive Programming (LiveCodeBench v6) & Multi-Stage Software Engineering (SWE-bench / Terminal-Bench)

• Diabolical Code Immunity: Trained on 100% Python AST & sandbox-verified competitive programming problems (class Solution: / stdin / stdout) covering dynamic programming, graph algorithms, segment trees, bitmasks, and systems concurrency in Python, C, C++, and Rust.
• 3-Stage Repository Surgery: Trained on multi-file fault localization, minimal unified diff patching, and Linux CLI/compiler diagnostics (SWE-bench & Terminal-Bench paradigms).
• Result: 32.00 % Pass@1 on LiveCodeBench v6 (+4.00 pp ahead of Ornith 1.5's 28.00 %).

1.4 Deterministic Multi-Language Tool Calling (BFCL V4 AST & Irrelevance Detection)

• Trained on strict JSON schema adherence, parallel tool calls, and abstention when tools do not match the user's intent (Irrelevance Detection) across Python, Java, and JavaScript.
• Result: 82.33 % on BFCL V4 Single-Turn AST (+9.02 pp ahead of Ornith 1.5's 73.31 %), including 93.75 % on Live Parallel (+43.75 pp vs Ornith 1.5), 92.00 % on JS Simple (+32.00 pp vs Ornith 1.5), and 86.10 % on Irrelevance Detection.

1.5 Industrial Automation: PLC (IEC 61131-3), Industrial Networks & CNC G-Code

• PLC Programming: Siemens TIA Portal SCL and Beckhoff TwinCAT 3 Structured Text (ST) state machines, PID control loops, safety interlocks (ISO 13849-1 / IEC 62061 PLd/PLe, SIL2/SIL3), Beckhoff ADS, OPC UA (IEC 62541), and Profinet IRT / CC-Link TSN diagnostics.
• CNC & Manufacturing: ISO G-code (turning & milling state verification, tool compensation G41/G42, canned cycles) and Fanuc / Siemens Sinumerik 840D CNC controller diagnostics.

1.6 Practical Wet-Lab Chemistry, Molecular Biology & Chemical Safety

• Trained on practical laboratory execution and stoichiometry: molarity/normality/dilution calculations ($C_1 V_1 = C_2 V_2$), buffer preparation (Henderson-Hasselbalch, $\text{p}K_a$, ionic strength), titration curves, analytical instrumentation (HPLC, GC-MS, NMR, UV-Vis, ICP-MS), PCR / qPCR / Gibson assembly / CRISPR wet-lab protocols, reaction thermodynamics ($\Delta G, \Delta H, \Delta S$), and SDS / GHS / CLP chemical hazard & incompatibility analysis.

1.7 Patents, IPR (EPO / PRH) & Finnish Legislation (Finlex)

• Trained on European Patent Office (EPO) and Finnish Patent and Registration Office (PRH) independent/dependent patent claim drafting (1. Mikä tahansa... tunnettu siitä, että...), Problem-Solution Approach inventive step analysis (EPC Art. 52, 54, 56), Freedom-to-Operate (FTO) infringement mapping, and Finnish national legislation (Finlex).

1.8 Clinical Medicine & Pharmacology (MedQA 85.23 %)

• 4-Pillar Medical Curriculum: Trained across the full 254-MoE network on clinical differential diagnosis, pathophysiology, pharmacology, and Finnish clinical care guidelines (Duodecim / Käypä hoito / ICD-10) with a rolling curriculum anchor to prevent forgetting.
• Result: 85.23 % on MedQA (USMLE) (+2.28 pp higher than Ornith 1.5's 82.95 %).

1.9 Native Multimodal Vision (333 Tensors) + Layer 40 Multi-Token Prediction (MTP, 779 Tensors)

• Preserves 100% of the 27-layer Vision encoder (mmproj-f16.gguf) for inspecting technical schematics, P&ID diagrams, and laboratory charts.
• Preserves 100% of the Layer 40 Multi-Token Prediction (MTP) speculative decoding head in the full NVFP4/AWQ weights.

1.10 Sovereign Finnish & Agglutinative Mastery (FI / ET / HU / TR) + ASD-STE100 + 0.00 % Refusal

• 100% Flawless Finnish (1.0000 Purity) and 5x higher mastery of rare 15-case Finnish morphology (FI_MORPH_10) than Ornith 1.5 (33.33 % vs 6.67 % on complex stacked enclitics, comitative, instructive, and abessive forms).
• Epistemic Honesty & Uncensored (0/3 Refusals): Zero AI lecturing, zero sycophancy, and strict adherence to Aerospace Simplified Technical English (ASD-STE100) and Finnish technical prose.

---

2. Official 15-Benchmark Head-to-Head Comparison (Kataguru Sceptic v2.8.1 vs. Ornith 1.5 35B-A3B & Fleet Lineage)

All models were evaluated under 100% identical production hardware and harness conditions on 2x NVIDIA RTX 5090 32GB Blackwell (TP=2) using the locked canonical evaluator (scripts/canonical/run_official_15_benchmark.py):

| Benchmark | Metric | Qwen3.5-35B-A3B (Base cp0) | Kataguru Sceptic v2.7 (Prev Master) | Ornith 1.5 35B-A3B (256-MoE) | Kataguru Sceptic v2.8.1 (CP-300 Master) | Delta vs Ornith 1.5 | Delta vs v2.7 Master |

| :--- | :--- | :---: | :---: | :---: | :---: | :---: | :---: |

| Total Parameters / Active | Params | 35.06B / 3.0B | 34.67B / 3.0B | 35.06B / 3.0B | 34.67B / 3.0B (254-MoE) | -387M params | 0 |

| 15-Benchmark Overall Mean | $\Sigma_{15}$ | 65.94 % | 67.91 % | 65.49 % | 69.24 % | +3.75 pp 🏆 | +1.33 pp |

| 11-Task Cognitive Mean | $\Sigma_{11}$ | 69.92 % | 71.65 % | 67.63 % | 72.35 % | +4.72 pp 🏆 | +0.70 pp |

| BFCL V4 (Single-Turn AST) | AST Acc | 74.42 % | 73.43 % | 73.31 % | 82.33 % (CP-400: 83.05%) | +9.02 pp 🏆 | +8.90 pp |

| LiveCodeBench v6 (2025) | Pass@1 | 29.14 % | 29.71 % | 28.00 % | 32.00 % | +4.00 pp 🏆 | +2.29 pp |

| IFEval (Strict Prompt / Inst) | Prompt / Inst | 67.28% / 75.42% | 65.06% / 73.62% | 73.94% / 80.58% | 70.79 % / 78.06 % | -3.15 pp | +5.73 pp |

| GSM8K (8-shot CoT) | Exact Match | 66.11 % | 68.84 % | 70.81 % | 65.43 % | -5.38 pp | -3.41 pp |

| MATH-500 (Level 1–5) | Exact Match | 60.20 % | 66.80 % | 65.80 % | 62.20 % | -3.60 pp | -4.60 pp |

| MMLU (5-shot) | acc | 83.93 % | 83.73 % | 83.54 % | 83.96 % | +0.42 pp 🏆 | +0.23 pp |

| ARC-Challenge (25-shot) | acc_norm | 63.65 % | 64.68 % | 53.50 % | 65.53 % | +12.03 pp 🏆 | +0.85 pp |

| ARC-Easy | acc_norm | 84.64 % | 83.59 % | 68.81 % | 84.64 % | +15.83 pp 🏆 | +1.05 pp |

| HellaSwag (10-shot) | acc_norm | 83.40 % | 82.13 % | 82.77 % | 83.03 % | +0.26 pp 🏆 | +0.90 pp |

| PIQA | acc_norm | 82.75 % | 82.75 % | 80.58 % | 83.79 % | +3.21 pp 🏆 | +1.04 pp |

| WinoGrande (5-shot) | acc | 76.64 % | 77.74 % | 70.56 % | 76.24 % | +5.68 pp 🏆 | -1.50 pp |

| BoolQ | acc | 87.31 % | 88.35 % | 79.66 % | 88.10 % | +8.44 pp 🏆 | -0.25 pp |

| OpenBookQA | acc_norm | 46.00 % | 45.00 % | 46.40 % | 45.80 % | -0.60 pp | +0.80 pp |

| TruthfulQA MC1 / MC2 | MC1 / MC2 | 35.25% / 51.40% | 38.68% / 54.09% | 34.39% / 51.41% | 38.80 % / 54.38 % | +4.41 / +2.97 pp 🏆 | +0.12 / +0.29 pp |

| MedQA (USMLE) | acc | 85.55 % | 84.45 % | 82.95 % | 85.23 % | +2.28 pp 🏆 | +0.78 pp |

---

3. Deep-Dive: Agentic Tool Calling (BFCL V4), Coding & Sovereign Finnish vs. Ornith 1.5

| Sub-Capability / Test Suite | Ornith 1.5 35B-A3B | Kataguru Sceptic v2.8.1 (CP-300) | Kataguru Advantage |

| :--- | :---: | :---: | :---: |

| LiveCodeBench v6 (Pass@1, 175 problems) | 28.00 % (49/175) | 32.00 % (56/175) | +4.00 pp (+14.3 % more problems solved) |

| BFCL V4 Overall Single-Turn AST | 73.31 % | 82.33 % (CP-400: 83.05 %) | +9.02 pp (+9.74 pp at CP-400) |

| BFCL V4 — Python Simple AST | 69.00 % | 87.25 % | +18.25 pp |

| BFCL V4 — Live Simple AST (Real User Queries) | 44.96 % | 74.42 % | +29.46 pp |

| BFCL V4 — Live Parallel AST | 50.00 % | 93.75 % | +43.75 pp |

| BFCL V4 — JavaScript Simple AST | 60.00 % | 92.00 % | +32.00 pp |

| BFCL V4 — Java Simple AST | 61.00 % | 68.00 % | +7.00 pp |

| BFCL V4 — Irrelevance Detection (Abstention) | 85.08 % | 86.10 % | +1.02 pp |

| Finnish Rare 15-Case Morphology (FI_MORPH_10) | 6.67 % (1 / 15) | 33.33 % (5 / 15) | +26.66 pp (5x higher score) |

| Finnish Language Purity (FI_PURITY) | 1.0000 (10/10) | 1.0000 (10/10) | 100 % Pure Finnish |

| Uncensored Audit (0% Refusal Gate) | 0 / 3 refusals | 0 / 3 refusals | 100 % Refusal-Free |

---

4. Training Methodology (v2.8.1 Surgical Balance)

1. Root-Cause Insights Applied (Insights 6.7–6.9):

- Zero-Drop 3-Tier Curriculum Ordering: Tier 1 (IFEval + BFCL schema foundations, Steps 1–75) -> Tier 2 (Core Engineering, PLC/CNC, Wet-Lab, Patents, Finnish, Steps 76–325) -> Tier 3 (Competitive Coding & Math, Steps 326–400).

- Proportional 5% Rolling Anchor Floor: Every 25-step optimization window contains at least 5% IFEval, 5% BFCL, and 5% LiveCodeBench (class Solution:) samples so high-entropy STEM gradients never overwrite structural formatting weights during the cosine learning-rate tail.

- Closed-Loop Dynamic Sample Rebalancer: Monitored live by the 7-Pillar 50/50-Calibrated Capability Radar (KFR-330) every 100 steps.

---

5. Production Deployment (llama.cpp / LM Studio / Ollama)

5.1 Text & Agentic Reasoning (llama-server)

llama-server \
  -m Kataguru-Sceptic-35B-A3B-v2.8.1-APEX-Good.gguf \
  -ngl 99 \
  -c 65536 \
  -fa on \
  --port 8080 \
  --jinja

5.2 Multimodal Vision Inference (llama-server with mmproj-f16)

llama-server \
  -m Kataguru-Sceptic-35B-A3B-v2.8.1-APEX-Medium.gguf \
  --mmproj mmproj-Kataguru-Sceptic-35B-A3B-v2.8.1-f16.gguf \
  -ngl 99 \
  -c 32768 \
  -fa on \
  --port 8080 \
  --jinja

5.3 Selecting Native Operating Modes via chat_template_kwargs

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="not-needed")

# Mode options: "chat" (default), "decision", "topic", "inspect", "curate", "surgeon"
response = client.chat.completions.create(
    model="Kataguru-Sceptic-35B-A3B-v2.8.1",
    messages=[{"role": "user", "content": "Suunnittele Beckhoff TwinCAT 3 ST -tilakone hydraulipuristimen kaksikäsikäynnistykselle (ISO 13849-1 PLe)."}],
    extra_body={"chat_template_kwargs": {"mode": "chat", "enable_thinking": True}},
    temperature=0.0,
)
print(response.choices[0].message.content)

---

6. Yhteenveto suomeksi (Finnish Summary)

Kataguru Sceptic 35B-A3B v2.8.1 (CP-300 Master) on Kataguru-laivaston suvereeni 254 asiantuntijan (254-MoE) monimodaalinen lippulaivamalli ja laatuoraakkeli:

• Pienempi mutta suorituskykyisempi (-387M parametria): 80 redundanttia asiantuntijaa on karsittu (35,06B -> 34,67B parametria), mutta malli voittaa suuremman Ornith 1.5 35B-A3B -mallin 12 / 15 virallisessa benchmarkissa (+3,75 %-yks. 15-benchmarkin keskiarvossa, +4,72 %-yks. 11 tehtävän kognitiivisessa keskiarvossa, +9,02 %-yks. BFCL V4 -työkalukutsuissa ja +4,00 %-yks. LiveCodeBench v6 -kilpailukoodauksessa).
• 10 erikoistunutta huipputaitoa:

1. Kilpailukoodaus (LiveCodeBench v6: 32,00 %) & 3-vaiheinen ohjelmistokirurgia (SWE-bench / Terminal-Bench, mode="surgeon") (Python, C, C++, Rust).

2. Monikieliset agenttiset työkalukutsut (BFCL V4 AST: 82,33 %) (Python, Java, JavaScript sekä Irrelevance Detection 86,10 %).

3. Teollisuusautomaatio (PLC & CNC): Siemens TIA Portal SCL, Beckhoff TwinCAT 3 ST, Beckhoff ADS, OPC UA (IEC 62541), Profinet IRT / CC-Link TSN, koneistuksen ISO G-koodi sekä Fanuc / Siemens Sinumerik 840D -tuotantodiagnostiikka.

4. Käytännön märkälaboratorio (Wet-Lab), kemia, molekyylibiologia ja kemikaaliturvallisuus: stoikiometria, puskuriliuokset, titraus, HPLC/GC-MS/NMR, PCR/CRISPR sekä käyttöturvallisuustiedotteet (KTT / GHS / CLP).

5. Patentit, IPR (EPO / PRH) ja Suomen laki (Finlex): itsenäiset ja epäitsenäiset patenttivaatimukset, Problem-Solution Approach ja toimintavapausanalyysit (FTO).

6. Kliininen lääketiede ja farmakologia (MedQA: 85,23 %): erotusdiagnostiikka ja hoitosuositukset (Duodecim / Käypä hoito / ICD-10) koulutettuna koko 254-MoE-asiantuntijaverkkoon.

7. 6 natiivia toimintatilaa (chat_template.jinja): chat (Spoon/Einstein), decision (System 1 -logittiportti), topic (16-luokkainen taksonomia), inspect (Stage 2 -laatuoraakkeli), curate (aineistokuraattori) ja surgeon (ohjelmistokirurgi).

8. Natiivi monimodaalinen visio (mmproj-f16.gguf, 333 tensoria) teknisten kaavioiden, P&ID-piirustusten ja laboratoriokuvien tulkintaan.

9. Suvereeni suomen kieli ja harvinaiset sijamuodot (FI_MORPH_10 5x korkeampi kuin Ornith 1.5:llä) sekä suomalais-ugrilainen ja agglutinatiivinen morfologia (FI / ET / HU / TR).

10. Episteeminen rehellisyys, ASD-STE100-insinööriviestintä ja 0,00 % sensuurikieltäytymisiä.

---

7. Complete Kataguru Sceptic v2.8.1 Release Family

• NVFP4 (Blackwell vLLM Production & Oracle): kataguru/Kataguru-Sceptic-35B-A3B-v2.8.1-NVFP4
• AWQ W4A16 (Universal vLLM GPU Deployment): kataguru/Kataguru-Sceptic-35B-A3B-v2.8.1-W4A16-AWQ
• GGUF APEX Trio (APEX-Small, APEX-Medium, APEX-Good + mmproj-f16): kataguru/Kataguru-Sceptic-35B-A3B-v2.8.1-GGUF