---

language:

• en

license: apache-2.0

base_model:

• kataguru/Kataguru-Industry-MVP-27B-v1.0-BF16

tags:

• gguf
• apex
• imatrix
• industrial
• automation
• plc
• beckhoff
• siemens
• rust
• c
• patents
• epo
• multimodal
• vision
• mtp
• uncensored
• reasoning

pipeline_tag: text-generation

---

Kataguru Industry-MVP-27B-v1.0 - GGUF APEX Series

> Sovereign industrial automation, systems programming (Rust/C), and patent engineering foundation model with dual-standard PLC support (Beckhoff TwinCAT & Siemens TIA Portal) and zero refusal (0.00%).

Official quantized GGUF release for Kataguru Industry-MVP-27B-v1.0 (Most Valuable Player) for local desktop execution in LM Studio, llama.cpp, Ollama, and edge inference servers.

---

Released GGUF Files

All quantized models were generated with llama.cpp using an importance matrix (imatrix) computed on Kataguru's high-entropy Finnish & technical calibration corpus.

| File | Size | Format | Description / Recommended Hardware |

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

| Kataguru-Industry-MVP-27B-v1.0-IQ4_XS.gguf | 15.0 GB | IQ4_XS | Flagship Quality. Best balance of precision and speed. Ideal for 16GB–24GB VRAM GPUs (RTX 4080/4090, RTX 5060 Ti 16GB, RTX 5080/5090) or 32GB Unified Memory Apple Silicon. |

| Kataguru-Industry-MVP-27B-v1.0-IQ3_S.gguf | 12.1 GB | IQ3_S | Balanced Efficiency. Fits comfortably on 16GB VRAM cards with plenty of context headroom. High precision retention. |

| Kataguru-Industry-MVP-27B-v1.0-IQ3_XXS.gguf | 10.8 GB | IQ3_XXS | Compact Lightweight. Extreme compression with minimal degradation. Runs on 12GB GPUs and lower-spec laptops. |

| mmproj-Kataguru-Industry-MVP-27B-v1.0-BF16.gguf | 889 MB | BF16 | Vision Projector. Multimodal image understanding and industrial visual inspection module. |

| mtp-Kataguru-Industry-MVP-27B-v1.0-Q8_0.gguf | 3.0 GB | Q8_0 | MTP Speculative Draft. Multi-Token Prediction speculative acceleration model. |

| mtp-Kataguru-Industry-MVP-27B-v1.0-Q4_K_M.gguf | 1.9 GB | Q4_K_M | MTP Speculative Draft (Compact). Low-memory Multi-Token Prediction draft model. |

| Kataguru-Industry-MVP-27B-v1.0.imatrix | 14 MB | GGUF | Pre-computed Importance Matrix for custom downstream quantization. |

---

Key Features & Capabilities

1. Native LM Studio & Minja Compatibility:

- The chat template has been verified and tested against the C++ Minja parser used in LM Studio and llama.cpp.

- Preserves native thinking/reasoning blocks in <think> ... </think> tags without template parsing crashes.

1. Dual-Standard Industrial Automation:

- Beckhoff TwinCAT 3 ST: IEC 61131-3 structured text, state machines, ADS communication, watchdog monitors (T#50MS).

- Siemens SCL (TIA Portal): Certified function blocks (FB), optimized memory layout, diagnostic alarm words (16#1001), and interlocks.

- Machining & CNC G-Code: Lathe and 5-axis mill state verification, constant surface speed (G96), feed rate compensation (G95), and modal cycle safety (G80/G00).

1. Systems Programming & Functional Safety:

- Rust & Modern C (C17/C23): Memory safety, zero-cost abstractions, non-blocking asynchronous event loops, and POSIX concurrency.

- ISO 13849-1 & IEC 62061: Formal calculation of Performance Level (PLr d/e, SIL 2/3), Mean Time to Dangerous Failure ($MTTF_D$), and Diagnostic Coverage ($DC_{avg}$).

1. Zero-Refusal & Anti-Sycophancy:

- 0.00% refusal rate on technical, scientific, and diagnostic inquiries.

- Epistemic honesty: Directly addresses technical problems, points out flawed assumptions, and refuses confabulation.

---

How to Run

LM Studio

1. Place the desired .gguf file (e.g. Kataguru-Industry-MVP-27B-v1.0-IQ4_XS.gguf) in your LM Studio models folder:

~/.lmstudio/models/kataguru/Kataguru-Industry-MVP-27B-v1.0-GGUF/

1. If vision support is needed, also download mmproj-Kataguru-Industry-MVP-27B-v1.0-BF16.gguf.
2. Load the model and chat!

> [!TIP]

> Recommended Sampling Parameters:

> - Temperature: 0.7

> - Top-P: 0.95

> - Repetition Penalty: 1.05 to 1.10 (prevents attractor repetition loops on technical enumerations)

llama.cpp CLI

# Text generation with IQ4_XS
llama-cli -m Kataguru-Industry-MVP-27B-v1.0-IQ4_XS.gguf \
  -p "<|im_start|>user\nSelitä lyhyesti TwinCAT 3 ADS-osoitteen rakenne.<|im_end|>\n<|im_start|>assistant\n" \
  -n 512 -c 8192 --temp 0.7 -ngl 99

# Multimodal inference with mmproj
llama-llava-cli -m Kataguru-Industry-MVP-27B-v1.0-IQ4_XS.gguf \
  --mmproj mmproj-Kataguru-Industry-MVP-27B-v1.0-BF16.gguf \
  --image electrical_schematic.png \
  -p "Analysoi tämän kytkentäkaavion turvapiirin hätäpysäytyskytkimet."

llama-server

llama-server -m Kataguru-Industry-MVP-27B-v1.0-IQ4_XS.gguf \
  --port 8080 -ngl 99 -c 16384

---

Lineage & Provenance

• Base Model: Kataguru-Industry-MVP-27B-v1.0-BF16
• Training Fleet: BIGIPA & AI-CORE (2x NVIDIA RTX 5090 32GB Blackwell)
• Dataset Engine: Kataguru Sovereign Sykli 1–3 Engineering Mixture
• Produced by: Kataguru AI Research Team