Kataguru-Sceptic-35B-A3B-v2.8.1-NVFP4 (Blackwell sm_120 Native FP4)
Kataguru-Sceptic-35B-A3B-v2.8.1-NVFP4 (CP-300 Master) is a bilingual (English & Finnish) 35-billion parameter (3B active parameters per token, 254 routed experts per layer, -387M fewer parameters than standard 256-expert Qwen3.5 MoE) multimodal Mixture-of-Experts model quantized to NVIDIA Blackwell NVFP4 (compressed-tensors, 25.44 GB) with synchronized w13 (gate_proj + up_proj) weight_global_scale across all 10,160 routed experts (0 scale mismatches).
1. Complete Skill Matrix & Specialized Capabilities
Unlike standard single-domain fine-tunes, Kataguru-Sceptic-35B-A3B-v2.8.1 combines general frontier reasoning with 10 specialized engineering, scientific, legal, medical, and data-factory capabilities in a single compact 254-expert MoE model:
| # | Capability Pillar | Trained Skills, Standards & Protocols | Verification & Architectural Mechanism |
|---|---|---|---|
| 1 | Compact 254-Expert MoE Architecture |
Pruned from 256 to 254 routed experts per layer (30,480 vs. 30,720 expert matrices), saving ~387M parameters (-1.30 GiB in BF16) with zero loss in active top_k=8 compute. |
10,160 routed experts + 40 shared experts (0 scale mismatches). |
| 2 | Native 6 Operating Modes (chat_template.jinja) |
chat ({REASON:spoon}, {REASON:einstein}, {REASON:deep}), decision (System 1 <20 ms 1-token logprob quality oracle), topic (System 1 learning-potential triage), inspect (JSON quality audit), curate (First-Principles synthesis), and surgeon ([MODE: SURGEON] / [TILA: KIRURGI]). |
Embedded directly in chat_template.jinja with both English and Finnish triggers. |
| 3 | Software Engineering, Competitive Coding & Multi-Turn SWE | Python, C/C++20, Rust, Java, and JavaScript algorithmic problem solving (class Solution:), plus 3-stage SWE-bench / Terminal-Bench workflows (Stage A: Bash Investigation $\rightarrow$ Stage B: Surgical git diff / replace_in_file $\rightarrow$ Stage C: pytest Verification). |
32.00 % Pass@1 on LiveCodeBench v6 (+4.00 pp over Ornith 1.5). All training code verified via PASSED_EXEC sandbox. |
| 4 | Multi-Language Agentic Tool Calling (BFCL V4) |
Single-turn, multi-turn, and parallel function calling (<tool_call>) across Python (92.75 %), Java (65.00 %), JavaScript (70.00 %), Multiple (95.50 %), and Parallel (84.50 %) tool schemas, plus negative tool-abstention (Irrelevance). |
82.33 % Overall Single-Turn AST (83.05 % at CP-400; +9.02 pp over Ornith 1.5). |
| 5 | Industrial Automation, PLC, CNC & Factory Diagnostics | IEC 61131-3 PLC: Siemens TIA Portal SCL & Beckhoff TwinCAT 3 Structured Text (ST) (CASE..OF state machines, R_TRIG, TON/TOF, safety interlocks). ISO 6983 CNC G-Code: Fanuc & Siemens Sinumerik 840D/Operate turning/milling cycles (G41/G42). Industrial Networks & SCADA: OPC UA (IEC 62541), Beckhoff ADS, Profinet IRT, and CC-Link TSN diagnostics. |
57.69 % on calibrated KFR-240 Industrial Automation radar (+7.69 pp over baseline). |
| 6 | Practical Wet-Lab Chemistry, Molecular Biology & Chemical Safety | Wet-Lab Protocols: Buffer preparation (Henderson–Hasselbalch), titration, fractional distillation, HPLC/GC-MS, NMR/IR/UV-Vis spectroscopy, PCR/qPCR primer design, gel electrophoresis, Michaelis–Menten enzyme kinetics, and CRISPR/Cas9 design. Chemical Safety & Materials: CLP hazard analysis, incompatible reagents, thermal runaway prevention, $\Delta G$ thermodynamics, Nernst electrochemistry, Fe-C phase diagrams, and lobby-free metabolic biochemistry (Canada 2019 Food Guide). | 84.64 % ARC-Easy (+15.83 pp vs. Ornith 1.5) & 65.53 % ARC-Challenge (+12.03 pp vs. Ornith 1.5). |
| 7 | Patents, IPR & Finnish Statutory Law | EPO & PRH Patent Engineering: Independent and dependent claim drafting (characterized in that... / tunnettu siitä, että...), EPO Problem-Solution Approach, inventive step evaluation, and Freedom-to-Operate (FTO) analysis. Finnish Law: Statutory interpretation (Finlex, KKO, KHO). |
Trained on certified EPO/PRH patent claim pairs and Finlex statutory corpus. |
| 8 | Clinical Medicine & Pharmacology (MedQA 85.23 %) |
USMLE clinical differential diagnosis, pharmacology, pathology, and Finnish clinical guidelines (Duodecim / Käypä hoito / ICD-10). | Trained across the full 254-MoE expert network via balanced rolling curriculum; 85.23 % MedQA (+2.28 pp over Ornith 1.5). |
| 9 | Multimodal Vision (visual.*) & Layer 40 MTP Speculation (mtp.*) |
Zero-Swap Vision: Reads HMI alarm screens, electrical schematics, and laboratory diagrams directly (333 vision tensors). Layer 40 MTP: Multi-Token Prediction speculative decoding (779 tensors, +12 % tok/s). |
100 % verified in verify_nvfp4_model_gate.py. |
| 10 | Sovereign Finnish, Epistemic Honesty, ASD-STE100 & 0 % Refusal | Rare Finnish case morphology (Kotus inflection & consonant gradation), anti-sycophancy (challenges false premises directly), ASD-STE100 Simplified Technical English, and SOMA/ARA orthogonalized 0.00 % refusal. |
5x higher score on rare Finnish morphology (FI_MORPH_10) than Ornith 1.5; 0 / 3 refusals (0.00 %). |
2. Empirical 15-Benchmark Comparison (Kataguru Sceptic v2.8.1 vs. Ornith 1.5 35B-A3B, v2.7 Master & Pre-SFT Base)
All models were evaluated on the exact same hardware and harness: 2x NVIDIA GeForce RTX 5090 32GB Blackwell (sm_120, TP=2, CUDAGraphs enabled, port 8003, lm-evaluation-harness 0.4.13 + LiveCodeBench v6 + BFCL V4). Complete evaluation logs for both Kataguru Sceptic v2.8.1 and Ornith 1.5 35B-A3B are included in eval/.
Despite having a smaller parameter footprint (254 vs. 256 routed experts per layer, -387M parameters), Kataguru-Sceptic-35B-A3B-v2.8.1 outperforms Ornith 1.5 35B-A3B on 12 out of 15 benchmarks, achieving:
- +3.75 pp higher 15-Benchmark Mean (
69.24 %vs.65.49 %) - +4.72 pp higher 11-Task Cognitive Mean (
72.35 %vs.67.63 %) - +4.00 pp higher LiveCodeBench v6 Pass@1 (
32.00 %vs.28.00 %) - +9.02 pp higher BFCL V4 Single-Turn AST (
82.33 %vs.73.31 %)
| # | Benchmark / Capability Metric | Kataguru Sceptic v2.8.1 (CP-300 Master) |
Ornith 1.5 35B-A3B (W4A16) |
$\Delta$ vs. Ornith 1.5 (pp) | Kataguru Sceptic v2.7 Master | $\Delta$ vs. v2.7 (pp) | Qwen3.5-35B-A3B Base (cp0) |
|---|---|---|---|---|---|---|---|
| $\Sigma_{15}$ | 15-Benchmark Unweighted Mean | 69.24 % (CP-400: 69.32 %) | 65.49 % | +3.75 pp | 67.91 % | +1.33 pp | — |
| $\Sigma_{11}$ | 11-Task Cognitive Mean (1–11) | 72.35 % | 67.63 % | +4.72 pp | 72.03 % | +0.32 pp | 61.26 % |
| 1 | PIQA (Physical Intuition, acc_norm) |
83.79 % | 80.58 % | +3.21 pp | 83.19 % | +0.60 pp | 81.72 % |
| 2 | WinoGrande (Contextual Reasoning, acc) |
75.77 % | 70.09 % | +5.68 pp | 75.85 % | -0.08 pp | 71.82 % |
| 3 | MedQA (USMLE 4-opts) (Clinical Medicine) | 85.23 % | 82.95 % | +2.28 pp | 85.39 % | -0.16 pp | 84.10 % |
| 4 | TruthfulQA MC2 (Epistemic Truthfulness) | 54.38 % | 51.41 % | +2.97 pp | 53.34 % | +1.04 pp | 52.89 % |
| 5 | TruthfulQA MC1 (Single-True Factuality) | 38.80 % | 34.39 % | +4.41 pp | 37.21 % | +1.59 pp | 34.39 % |
| 6 | ARC-Easy (Standard STEM, acc_norm) |
84.64 % | 68.81 % | +15.83 pp | 84.26 % | +0.38 pp | 74.83 % |
| 7 | HellaSwag (Linguistic Coherence) | 83.03 % | 80.94 % | +2.09 pp | 82.53 % | +0.50 pp | 83.30 % |
| 8 | OpenBookQA (Scientific Application) | 45.80 % | 44.40 % | +1.40 pp | 44.80 % | +1.00 pp | 46.00 % |
| 9 | ARC-Challenge (Advanced STEM, acc_norm) |
65.53 % | 53.50 % | +12.03 pp | 63.74 % | +1.79 pp | 57.94 % |
| 10 | BoolQ (Natural Language Logic, acc) |
90.00 % | 81.56 % | +8.44 pp | 90.55 % | -0.55 pp | 85.78 % |
| 11 | GSM8K Math (CoT) (strict / flex) |
88.93 % / 90.22 % | 95.30 % / 95.38 % | -6.37 pp | 91.51 % / 91.51 % | -2.58 pp | 1.06 % / 12.36 % |
| 12 | GPQA Diamond (198 PhD STEM items) | 57.58 % (CP-400: 61.11 %) | 62.63 % | -5.05 pp (-1.52 vs CP400) | 58.08 % | -0.50 pp (+3.03 CP400) | — |
| 13 | IFEval Strict (prompt / inst) |
70.79 % / 78.06 % | 74.49 % / 80.22 % | -3.70 pp / -2.16 pp | 65.06 % / 73.62 % | +5.73 pp / +4.44 pp | — |
| 14 | LiveCodeBench v6 (Pass@1, 175 problems) |
32.00 % | 28.00 % | +4.00 pp | 29.71 % | +2.29 pp | — |
| 15 | BFCL V4 Single-Turn AST (Overall) | 82.33 % (CP-400: 83.05 %) | 73.31 % | +9.02 pp | 73.43 % | +8.90 pp | — |
2.1 Coding, Agentic Tool Calling (BFCL V4) & Finnish Morphology Breakdown (Kataguru Sceptic v2.8.1 vs. Ornith 1.5)
| Sub-Benchmark / Gate Probe | Kataguru Sceptic v2.8.1 (CP-300 Master) |
Ornith 1.5 35B-A3B | $\Delta$ (Sceptic vs. Ornith) |
|---|---|---|---|
LiveCodeBench v6 (Pass@1) |
32.00 % | 28.00 % | +4.00 pp |
| BFCL V4 Non-Live AST Summary | 85.23 % (CP-400: 86.31 %) | 76.83 % | +8.40 pp |
| BFCL V4 Live AST Summary | 79.42 % (CP-400: 79.79 %) | 69.80 % | +9.62 pp |
| BFCL Python Simple AST | 92.75 % (CP-400: 93.75 %) | 74.50 % | +18.25 pp |
| BFCL Live Simple AST | 82.95 % (CP-400: 84.11 %) | 53.49 % | +29.46 pp |
| BFCL Multiple Functions AST | 95.50 % | 86.50 % | +9.00 pp |
| BFCL Parallel Functions AST | 84.50 % (CP-400: 87.00 %) | 78.00 % | +6.50 pp |
| BFCL Live Parallel AST | 68.75 % | 25.00 % | +43.75 pp |
| BFCL JavaScript Simple AST | 70.00 % (CP-400: 72.00 %) | 38.00 % | +32.00 pp |
| BFCL Java Simple AST | 65.00 % (CP-400: 66.00 %) | 64.00 % | +1.00 pp |
| BFCL Irrelevance Detection | 46.87 % (Non-Live: 58.33 %) | 71.70 % | -24.83 pp |
Finnish Rare Case Morphology (FI_MORPH_10) |
5 / 15 | 1 / 15 | 5x higher accuracy on rare Finnish inflection |
Finnish Engineering Register (FI_SYNTAX_REGISTER) |
4 / 5 | 3 / 5 | +20.00 pp |
Uncensored Refusal Gate (3 probes) |
0 / 3 refusals (0.00 %) |
0 / 3 refusals (0.00 %) |
Tied (100 % non-refusal) |
3. Native 6 Operating Modes (chat_template.jinja)
The embedded chat_template.jinja recognizes both English and Finnish trigger prefixes (case-insensitive) at the start of the user prompt, or via chat_template_kwargs={"sceptic_mode": "..."}:
Mode (sceptic_mode) |
English Text Prefix | Finnish Text Prefix | Behavior & Description |
|---|---|---|---|
chat (default) |
(none) | (ei etuliitettä) | General engineering, coding, wet-lab, legal, and reasoning assistant. Supports {REASON:spoon}, {REASON:einstein}, {REASON:deep}/{REASON:xhigh}, {REASON:high}, {REASON:medium}, {REASON:low}, {REASON:none}/{REASON:off}. |
decision |
[MODE: DECISION] or [SYSTEM 1 DECISION] |
[LAATUORAAKKELI: SYSTEEMI 1] |
System 1 logprob quality oracle (0–9, < 20 ms, 0 output tokens). Appends single-digit scoring prompt and disables <think>. |
topic |
[MODE: TOPIC] or [SYSTEM 1 TOPIC] |
[LAATUORAAKKELI: AIHE] |
System 1 topic potential oracle (0–9). Scores prompt learning potential to separate STEM diamonds ($E[\text{topic}] \ge 6.0$) from trivial queries. |
inspect |
[MODE: INSPECT] or [QUALITY INSPECTOR] |
[TILA: TARKISTUS] |
System 2 Verdict-First JSON quality inspector (`{"verdict": "ACCEPT" |
curate |
[MODE: CURATE] or [DATA CURATOR] |
[TILA: KURAATTORI] or [TILA: KOULUTUS] |
First-Principles dataset synthesis and audit mode. |
surgeon |
[MODE: SURGEON] or [DATA SURGEON] |
[TILA: KIRURGI] |
Autonomous Data Surgeon: rewrites failing derivations and broken Python code blocks from first principles so they pass deterministic sandbox execution. |
4. Production Deployment (vLLM on 2x RTX 5090 Blackwell TP=2)
vllm serve kataguru/Kataguru-Sceptic-35B-A3B-v2.8.1-NVFP4 \
--served-model-name kataguru/Kataguru-Sceptic-35B-A3B-v2.8.1-NVFP4 \
--tensor-parallel-size 2 \
--quantization compressed-tensors \
--kv-cache-dtype fp8 \
--gpu-memory-utilization 0.85 \
--max-model-len 262144 \
--enable-chunked-prefill \
--speculative-config '{"method":"mtp","num_speculative_tokens":3}' \
--limit-mm-per-prompt '{"image":10}' \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder \
--trust-remote-code
Suomenkielinen Yhteenveto (Kataguru-Sceptic-35B-A3B-v2.8.1)
Kataguru-Sceptic-35B-A3B-v2.8.1 (CP-300 Master) on suomen- ja englanninkielinen 35 miljardin parametrin (3B aktiivista parametria / token, 254 reititettyä asiantuntijaa / kerros, -387M parametria kevyempi kuin vakio 256-MoE) multimodaalinen Mixture-of-Experts -lippulaivamalli.
Koulutetut erikoistaidot
- Kuusi sisäänrakennettua toimintatilaa (
chat_template.jinja): Yleinen päättely ({REASON:spoon},{REASON:einstein}), System 1 pikapäätös (decision&topic,< 20 ms, 0 ulostulotokenia), Quality Inspector (inspect), Kuraattori (curate) ja Autonominen Datakirurgi ([TILA: KIRURGI]). - Ohjelmistotuotanto, Algoritmit & Monivuoroinen SWE-korjaus: Python, C/C++20, Rust, Java ja JavaScript (
LiveCodeBench v6 Pass@1: 32,00 %) sekä 3-vaiheisetSWE-bench/Terminal-Bench-korjausketjut. - Monikieliset Agenttityökalut (
BFCL V4: 82,33 %): Python (92,75 %), Java (65,00 %), JavaScript (70,00 %), Multiple (95,50 %) ja Parallel (84,50 %)<tool_call>-kutsut. - Teollisuusautomaatio, PLC, CNC & Tehdasdiagnostiikka: Siemens TIA Portal SCL, Beckhoff TwinCAT 3 ST (IEC 61131-3), Fanuc / Sinumerik 840D G-koodi (ISO 6983), OPC UA (IEC 62541), Beckhoff ADS sekä Profinet IRT / CC-Link TSN.
- Käytännön Märkälaboratorio (Wet-Lab), Kemia, Molekyylibiologia & Kemikaaliturvallisuus: Puskuriliuokset, titraus, tislaus, HPLC/GC-MS, NMR/IR, PCR/qPCR, CRISPR/Cas9, entsyymikinetiikka, CLP-kemikaaliturvallisuus, reaktiivisuus, termodynamiikka ($\Delta G$, Nernst, Fe-C-faasidiagrammit) ja riippumaton metaboliabiokemia.
- Patenttioppi (EPO/PRH), IPR & Suomen Lainsäädäntö: Patenttivaatimusten (
tunnettu siitä, että...) laadinta, Problem-Solution Approach, FTO-analyysi ja suomalainen lainsäädäntö (Finlex,KKO/KHO). - Kliininen Lääketiede (
MedQA: 85,23 %): Erotusdiagnostiikka ja farmakologia (Duodecim / Käypä hoito / ICD-10) koulutettuna koko254-MoE-asiantuntijaverkkoon. - Multimodaalinen Visio (
333tensoria) + Layer 40 MTP (779tensoria) + Suvereeni Suomi & 0 % Refusal.