Kataguru Industry-MVP-27B-v1.1 (AWQ W4A16)
Sovereign Industrial Automation, Systems Engineering (Rust/C), and Patent Jurisprudence Foundation Model with Dual-Standard PLC Support (Beckhoff TwinCAT & Siemens TIA Portal), Native Multimodal Vision, MTP Speculation, Zero Refusal (0.00%), Vahvistettu Suomen Kielituki (Confirmed Native Finnish Support), and AWQ W4A16 Tensor-Parallel Optimization.
🇫🇮 Vahvistettu Suomen Kielituki (Confirmed & Verified Native Finnish Support)
Kataguru Industry-MVP-27B-v1.1 on varustettu täydellä, empiirisesti todennetulla ja laitteistotestatulla suomen kielen tuella (Confirmed & Verified Native Finnish Language Support):
- Kotus-Morfologinen Stressitesti (100 % läpäisy, 15/15): Todennuskoe suomen kielen vaativimmille kielenpiirteille läpäisty täysin virheettömästi (kpt-astevaihtelu, monitavuiset yhdyssanat, possessiivisuffiksit, liitepartikkelit -kin/-kaan sekä epäsäännölliset taivutusvartalot).
- Teollisuuden ja Insinööritieteen Suomi: Hallitsee suomenkielisen automaatio-, sähkö- ja konepajasanaston (Beckhoff TwinCAT, Siemens TIA Portal, CNC G-koodi, ISO 13849-1 -koneturvallisuus) ilman käännöskukkasia tai kankeita anglismeja.
- Patentti- ja IPR-Suomi (PRH): Tuottaa muodollisesti päteviä suomenkielisiä patenttivaatimuksia (PRH: itsenäiset ja epäitsenäiset vaatimukset, tunnusmerkkiosa ja laiteosa) sekä EPO-problem-solution -päättelyketjuja.
- Suomenkielinen Kalibrointi: AWQ-skaalaimet on optimoitu suomalaisella teknisellä corpora-aineistolla, mikä takaa, ettei 4-bittinen painopakkaus heikennä suomen kielen morfosyntaktista tarkkuutta.
Change Log: What did Kataguru change in v1.1?
| Specification | Details |
|---|---|
| Base Architecture | Qwen3.8-27B Dense (Qwen3_5ForConditionalGeneration with active Vision & MTP heads) |
| Kataguru Modifications | Full 3-Cycle Sovereign Engineering Curriculum (46,000 samples) + Sykli 2.5 Sovereign Finnish Alignment (763 steps) + AMK CP619 Stacking + SOMA/ARA Zero-Refusal Orthogonalization + 150-step Lightweight Anchor SFT |
| What was NOT changed | Native architecture, core tensor shapes, full uncompressed BF16 vision tower (333 weights), and MTP speculative decoding head (15 weights) |
| Primary Use Cases | Industrial automation engineering (TwinCAT 3 ST, Siemens SCL, CNC G-code), mission-critical systems programming (Rust/C), patent & FTO analysis, multimodal technical inspection, uncensored first-principles technical diagnostics |
| Target Deployments | vLLM cluster inference (AWQ W4A16 & Blackwell NVFP4) and desktop workstations (GGUF APEX trio: IQ3_XXS, IQ3_S, IQ4_XS) |
| Tested Hardware | 2x NVIDIA RTX 5090 32GB Blackwell (TP=2), RTX 5060 Ti 16GB, RTX 4090 24GB |
| Refusal Rate | 0.00 % (Zero refusals across biochemical synthesis, systems exploitation, and industrial chemistry) |
| Finnish Language Mastery | 100 % pass on 15 Kotus morphological stress cases (nominal compounding, consonant gradation, non-linear agglutination) |
Official 13-Benchmark Verification (2x RTX 5090 Blackwell TP=2, CUDAGraphs)
All evaluations are conducted under strict Zero-Mocking & Hardware Verification principles on production hardware (2x NVIDIA RTX 5090 32GB Blackwell, Tensor Parallelism TP=2, port 8003) using identical production parameters.
| Benchmark Suite | Metric / Standard | Industry-MVP-27B-v1.0 | Industry-MVP-27B-v1.1 | Delta ($\Delta$) | Status |
|---|---|---|---|---|---|
| IFEval | Strict Prompt Accuracy | 70.43 % | 70.98 % | +0.55 % | Upgrade |
| IFEval | Strict Instruction Accuracy | 79.26 % | 79.50 % | +0.24 % | Upgrade |
| GSM8K | Math CoT (Strict Exact Match) | 89.01 % | 89.54 % | +0.53 % | Upgrade |
| GSM8K | Math CoT (Flexible Match) | 90.14 % | 90.45 % | +0.31 % | Upgrade |
| GPQA Diamond | High-Difficulty Reasoning (Zero-Shot) | 60.10 % | 62.12 % (123/198) | +2.02 % | Fleet Record |
| WinoGrande | Common Sense Reasoning | 77.58 % | 78.45 % | +0.87 % | All-Time 27B Record |
| MedQA | USMLE Medical Board Exam (4-options) | 80.28 % | 80.52 % | +0.24 % | Robust Preservation |
| BoolQ | Boolean Question Answering | 90.06 % | 90.31 % | +0.25 % | Pass |
| ARC-Easy | Science Question Answering | 88.35 % | 88.47 % | +0.12 % | Pass |
| ARC-Challenge | Complex Scientific Reasoning | 67.24 % | 67.49 % | +0.25 % | Pass |
| HellaSwag | Multi-Choice Sentence Completion | 82.85 % | 82.90 % | +0.05 % | Pass |
| OpenBookQA | Open Book Scientific Reasoning | 45.20 % | 45.40 % | +0.20 % | Pass |
| PIQA | Physical Interaction QA | 82.15 % | 82.21 % | +0.06 % | Pass |
| TruthfulQA | MC2 / MC1 Truthfulness | 54.32% / 36.31% | 54.49 % / 36.47 % | +0.17 % | Pass |
| $\Sigma_{11}$ Standard Average | Fleet Macro Benchmark Average | 72.18 % | 72.39 % | +0.21 % | Fleet Record |
Architectural Moat & Engineering Pillars
1. Dual-Standard Industrial Automation
- Beckhoff TwinCAT 3 ST (IEC 61131-3): Full support for structured text state machines, ADS client-server protocols, cyclic watchdog timers (
T#50MS), and dual-channel sensor discrepancy supervision. - Siemens SCL (TIA Portal): Optimized function blocks (
FB), static memory encapsulation, diagnostic alarm words (16#1001), and safe axis interlocks. - CNC Machining & G-Code: Lathe and 5-axis state validation, constant surface speed (
G96), feed per revolution (G95), and modal cycle safety (G80/G00). - Functional Safety (ISO 13849-1): Performance Level calculation ($PL_r = e$), Category 4 dual-channel architecture, $MTTF_d$, and diagnostic coverage ($DC_{avg}$).
2. Systems Programming in Rust & Embedded C
- Rust Systems Engineering: Zero-allocation lock-free ring buffers, explicit
Acquire/Releasememory barriers, CAS concurrency, safe FFI bindings, and ownership modeling. - Embedded C & x86_64: Memory-mapped I/O (MMIO), DMA descriptor ring queues, CPU cache invalidation, and interrupt service routines.
3. European & Finnish Patent Jurisprudence (EPO & PRH)
- Rigorous execution of the EPO Problem-Solution Approach: identifying closest prior art, technical effect, objective technical problem, and non-obviousness under EPC Art. 56.
- Freedom-to-Operate (FTO) claim chart cross-examinations and Finnish PRH patent claim structure synthesis.
4. 100% Uncompromised Multimodality (Vision & MTP)
- Vision Tower: Preserves all 333 visual weights (
model.visual.*) in uncompressed high precision. Capable of technical diagram reading, schematic analysis, and high-detail scene comprehension. - MTP Speculative Decoding: Retains all 15 Multi-Token Prediction weights (
mtp.*) for multi-token speculative acceleration.
5. Epistemic Honesty & Zero Refusal (0.00%)
- Refusal direction removal via SOMA/ARA activation orthogonalization across 66 target matrices in layers 16–48.
- 0.00% refusal across biochemical synthesis, systems exploitation, and industrial chemistry.
Quickstart & Deployment
vLLM (Production Cluster Inference)
vllm serve kataguru/Kataguru-Industry-MVP-27B-v1.1-AWQ \
--tensor-parallel-size 2 \
--quantization awq \
--dtype float16 \
--max-model-len 32768 \
--port 8003
License & Attribution
Released under the Apache 2.0 license. Developed and verified by Kataguru Sovereign AI Engineering.