---

language:

• en
• fi

license: apache-2.0

base_model:

• kataguru/Kataguru-Industry-MVP-27B-v1.1-BF16

tags:

• finnish
• suomi
• sovereign-finnish
• confirmed-finnish-support
• industrial
• automation
• plc
• beckhoff
• siemens
• rust
• c
• patents
• epo
• prh
• iso-13849-1
• multimodal
• vision
• mtp
• uncensored
• reasoning
• gguf
• lmstudio
• 16gb-vram
• rtx-5060-ti
• dflash2
• tqserve

pipeline_tag: text-generation

---

Kataguru Industry-MVP-27B-v1.1 (GGUF APEX Suite & 16 GB Turnkey Appliance)

> 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/DFlash2 Speculation, Zero Refusal (0.00%), Confirmed Native Finnish Support, imatrix-optimized GGUF APEX Quantizations, and a 1-Command 16 GB VRAM Turnkey Appliance (98.42 tok/s & 64k context on a single RTX 5060 Ti 16 GB).

---

⚡ 16 GB VRAM Turnkey Appliance (tqserve + DFlash2: 98.42 tok/s & 64k Context)

This repository includes a self-contained 1-Command 16 GB VRAM Turnkey Appliance that runs the 27-billion-parameter Kataguru-Industry-MVP-27B-v1.1-IQ3_XXS.gguf model on a single 16 GB consumer GPU (e.g., NVIDIA GeForce RTX 5060 Ti 16 GB) with a full 65,536-token (64k) context window at 67.5–98.4 tok/s decode speed.

The appliance bundles:

• Kataguru-Industry-MVP-27B-v1.1-IQ3_XXS.gguf (10.35 GiB, 3.31 BPW Finnish-calibrated imatrix quant)
• Qwen3.8-27B-DFlash2-TQ3.gguf (750 MB DFlash2 block-diffusion speculative drafter)
• appliance_16gb/bin/tqserve (Pre-compiled C++/CUDA Blackwell SM120 inference engine from xhinker/azllm / turbo-tan/tqllm with TQ3_0 14 KB/token KV cache)
• appliance_16gb/config/kataguru_chat_template.jinja (Official Kataguru Tri-Mode Jinja chat template)

1. One-Command Install & Launch (Smart Downloader)

Run the following command on any Linux x86_64 workstation with a 16 GB+ NVIDIA GPU. The smart launcher automatically downloads only the required 16 GB appliance files (skipping larger quants) and starts an OpenAI-compatible /v1/chat/completions server on port 8086:

curl -sSL https://huggingface.co/kataguru/Kataguru-Industry-MVP-27B-v1.1-GGUF/resolve/main/start_16gb_appliance.sh | bash -s -- --gpu 0 --port 8086

Or run locally after downloading start_16gb_appliance.sh:

chmod +x start_16gb_appliance.sh
./start_16gb_appliance.sh --gpu 0 --port 8086 --ctx 65536

2. OpenAI API Usage Example (http://127.0.0.1:8086/v1/chat/completions)

curl http://127.0.0.1:8086/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kataguru-27b",
    "messages": [
      {"role": "user", "content": "Write a Beckhoff TwinCAT 3 Structured Text (IEC 61131-3) function block FB_HydraulicPressSafety for two-hand synchronous control (max 500 ms discrepancy) and light curtain monitoring."}
    ],
    "max_tokens": 1024,
    "temperature": 0.0,
    "chat_template_kwargs": {"enable_thinking": false}
  }'

3. Hardware-Verified Benchmark: 1x NVIDIA GeForce RTX 5060 Ti 16 GB (--context 65536 --kv tq3)

| Benchmark / Metric | Measured Result on 1x RTX 5060 Ti 16 GB (180.16 W) |

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

| Model & Precision | Kataguru-Industry-MVP-27B-v1.1-IQ3_XXS.gguf (10.35 GiB, 3.31 BPW imatrix) |

| Speculative Drafter | Qwen3.8-27B-DFlash2-TQ3.gguf (750 MB, DFlash2 block-diffusion) |

| Allocated Context Window (--kv tq3) | 65,536 tokens (64k) (14.0 KB/token TQ3_0 KV cache) |

| VRAM Footprint (64k ctx + 8,845-token prompt) | 15,280 / 16,311 MiB (>1,030 MiB free VRAM margin) |

| Prefill Throughput (8,845-token prompt) | 1,286.5 tok/s |

| Decode Speed: Beckhoff TwinCAT 3 ST (IEC 61131-3) | 98.42 tok/s (41.5% draft acceptance, 91.9 tok/s wall-clock) |

| Decode Speed: Long-Context Summarization (at 9k ctx) | 67.52 tok/s |

---

🇫🇮 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.
• Suomalainen Aktivaatiomatriisi (imatrix): Kaikki GGUF-kvantit on optimoitu suomalaisella teknisellä corpora-aineistolla (Kataguru-Industry-MVP-27B-v1.1.imatrix), jolloin kvantisointi ei riko suomen kielen rikasta agglutinoivaa morfematiikkaa.

---

Release Formats & Quantization Hierarchy

| File Name | Format / Quant | Size | Recommended Use Case | Hardware Fit |

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

| Kataguru-Industry-MVP-27B-v1.1-IQ3_XXS.gguf | APEX-Small | ~10.9 GB | Maximum efficiency & 16 GB Turnkey Appliance (64k ctx) | 12GB–16GB VRAM (e.g. RTX 4060 Ti / 5060 Ti 16GB) |

| Kataguru-Industry-MVP-27B-v1.1-IQ3_S.gguf | APEX-Medium | ~12.2 GB | Optimal balance between speed and precision | 16GB VRAM (e.g. RTX 4070 / 5060 Ti 16GB) |

| Kataguru-Industry-MVP-27B-v1.1-IQ4_XS.gguf | APEX-Good | ~15.1 GB | High fidelity, production desktop standard | 24GB–32GB VRAM (e.g. RTX 4090 / 5090) |

| Qwen3.8-27B-DFlash2-TQ3.gguf | DFlash2 Speculative Drafter | ~750 MB | Block-diffusion drafter for tqserve (up to 98.4 tok/s on 16GB GPU) | Loaded via start_16gb_appliance.sh |

| start_16gb_appliance.sh | 16 GB Turnkey Launcher | ~5 KB | Smart 1-command downloader & OpenAI server launcher for 16 GB GPUs | Linux x86_64 + CUDA 12.8 (appliance_16gb/) |

| mmproj-Kataguru-Industry-MVP-27B-v1.1-BF16.gguf | Vision Projector | ~889 MB | Required for multimodal image/diagram reasoning | Loaded alongside any base GGUF quant |

| mtp-Kataguru-Industry-MVP-27B-v1.1-Q4_K_M.gguf | MTP Speculative Head | ~1.9 GB | Speculative decoding acceleration (Q4) | Compatible llama.cpp engines |

| mtp-Kataguru-Industry-MVP-27B-v1.1-Q8_0.gguf | MTP Speculative Head | ~3.0 GB | Speculative decoding acceleration (Q8) | Compatible llama.cpp engines |

| Kataguru-Industry-MVP-27B-v1.1.imatrix | Calibration Matrix | ~14 MB | Finnish domain activation matrix | Used during APEX quantization |

All quantizations are calibrated using native Blackwell llama-imatrix with domain-specific Finnish technical corpora to preserve 100% Finnish morphological integrity and zero-shot technical reasoning.

---

Official 13-Benchmark Verification (2x RTX 5090 Blackwell TP=2, CUDAGraphs)

| 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 |

---

Deployment in LM Studio

1. Download your preferred APEX quant (e.g. Kataguru-Industry-MVP-27B-v1.1-IQ4_XS.gguf) and the vision projector mmproj-Kataguru-Industry-MVP-27B-v1.1-BF16.gguf.
2. Move the files to your LM Studio model folder:

```bash

mkdir -p ~/.lmstudio/models/kataguru/Kataguru-Industry-MVP-27B-v1.1-GGUF/

mv Kataguru-Industry-MVP-27B-v1.1-IQ4_XS.gguf ~/.lmstudio/models/kataguru/Kataguru-Industry-MVP-27B-v1.1-GGUF/

mv mmproj-Kataguru-Industry-MVP-27B-v1.1-BF16.gguf ~/.lmstudio/models/kataguru/Kataguru-Industry-MVP-27B-v1.1-GGUF/

```

1. In LM Studio, load the model with:

- GPU Offload: Max / 100%

- Context Length: 32,768 (or 131,072 depending on VRAM)

- Temperature: 0.2

- Vision: Select mmproj-Kataguru-Industry-MVP-27B-v1.1-BF16.gguf

---

Deployment via llama.cpp / llama-cli

llama-cli \
    -m Kataguru-Industry-MVP-27B-v1.1-IQ4_XS.gguf \
    --mmproj mmproj-Kataguru-Industry-MVP-27B-v1.1-BF16.gguf \
    -ngl 99 \
    --temp 0.2 \
    -c 32768 \
    -p "<|im_start|>system\nOlet Kataguru Industry MVP 27B, teollisuusautomaation ja insinööritieteen suvereeni erikoismalli.<|im_end|>\n<|im_start|>user\nAnalysoi oheisen kytkentäkaavion turvalogiikan toteutus ISO 13849-1 mukaisesti.<|im_end|>\n<|im_start|>assistant\n"

---

License & Attribution

Released under the Apache 2.0 license. Developed and verified by Kataguru Sovereign AI Engineering.