kataguru/Industry-Coder-27B-v1.0-GGUF (4 APEX/IQ Quants + Vision + MTP)

Kataguru Industry-Coder-27B-v1.0 is a 27-billion-parameter dense flagship model trained via a 4-cycle closed-loop curriculum (Cycle 1..4 + Tempering Elite Bank / Karkaisupankki) for industrial software engineering, algorithmic problem solving (LiveCodeBench v6), agentic tool calling (BFCL V4), industrial automation (PLC/CNC: Beckhoff TwinCAT 3 ST & Siemens SCL), European/Finnish patent engineering (EPO/PRH), and doctoral-level STEM reasoning (GPQA Diamond & GSM8K).

🇫🇮 Suomenkielinen dokumentaatio löytyy englanninkielisen osion alta.

🇬🇧 English Documentation

1. Summary of Kataguru Modifications

Field	Description / Value
Base model	`Qwen 3.8 27B` (64 layers: Gated DeltaNet + Full Attention + 333 Vision tensors + 15 MTP speculative tensors)
Kataguru modifications	4-cycle closed-loop Merge-and-Reset curriculum (`Cycle 1 Foundation` $\to$ `Cycle 2 Upper Secondary` $\to$ `Cycle 3 Applied Engineering` $\to$ `Cycle 4 Tempering Elite Bank`), `r=16 qr_frozen_a` orthonormal subspace training, `kataguru_trisection` LR schedule, SOMA/ARA `0.00%` refusal alignment, and Kataguru Quad-Mode Chat Template
Language quality (EN & FI)	Native-level English and 100% flawless Finnish: natural, idiomatic Finnish across software engineering, industrial automation, patents, and STEM without morphological errors or calques (calibrated with bilingual `Industry-Coder-27B-v1.0.imatrix`)
Quantization (4 APEX/IQ models)	4 imatrix-calibrated GGUF models: `IQ3_XXS` (smallest, ~11 GB), `IQ3_S` (~12 GB), `IQ4_XS` (~15 GB), `IQ4_NL` (~15 GB) + `mmproj-f16` (vision) + `mtp-Q8_0` / `mtp-Q4_K_M` (MTP speculative decoding)
Weights size (Disk)	`10.6 GB` (`IQ3_XXS`) to `15.3 GB` (`IQ4_NL`) + `889 MB` (`mmproj`)
VRAM requirement (GGUF)	12 GB GPU (`IQ3_XXS` with moderate context), 16 GB GPU (`IQ3_XXS` / `IQ3_S` / `IQ4_XS`), 24 GB+ GPU (`IQ4_NL`)
What was NOT modified	64-layer hybrid backbone architecture, 333 vision tensors (`mmproj`), 15 MTP tensors, and 262,144-token native context window
Primary use cases	Production software engineering (Python, Rust, C, Java, JS), agentic `<tool_call>` workflows, PLC/CNC automation, patent claims, and STEM reasoning
Recommended runtime	LM Studio / `llama.cpp` (`b11064+`) with 100% GPU Offload
Tested environment	2x NVIDIA RTX 5090 32GB Blackwell (TP=2, vLLM v0.29.0-cu129, CUDAGraphs) & `llama.cpp b11064`

2. Included Files (4 APEX/IQ GGUF Quants + Vision + MTP)

Filename	Precision / Profile	Size	Recommended Hardware & Use Case
`Industry-Coder-27B-v1.0-IQ3_XXS.gguf`	`IQ3_XXS` (`3.06 bpw`, APEX-Smallest)	11 GB	Smallest 12 GB / 16 GB GPU quant: Fits on 12 GB GPUs with moderate context and leaves >5 GB KV cache headroom on 16 GB GPUs.
`Industry-Coder-27B-v1.0-IQ3_S.gguf`	`IQ3_S` (`3.44 bpw`, APEX-Small/Med)	12 GB	Fast 16 GB GPU quant: High decode throughput and generous KV cache headroom on 16 GB GPUs (e.g., RTX 5060 Ti 16GB / RTX 4080).
`Industry-Coder-27B-v1.0-IQ4_XS.gguf`	`IQ4_XS` (`4.25 bpw`, APEX-Medium/Good)	15 GB	16–24 GB Sweet Spot: High-accuracy 4-bit non-linear imatrix quant.
`Industry-Coder-27B-v1.0-IQ4_NL.gguf`	`IQ4_NL` (`4.50 bpw`, APEX-Max)	15 GB	16–24 GB+ Max Quality: Highest-precision 4-bit non-linear GGUF quant.
`mmproj-Industry-Coder-27B-v1.0-f16.gguf`	`F16` Vision	889 MB	Multimodal vision projector for schematics, diagrams, and technical documents.
`mtp-Industry-Coder-27B-v1.0-Q8_0.gguf`	`Q8_0` MTP Draft	3.0 GB	Layer 64 (`blk.64.nextn`) Multi-Token Prediction speculative head (high precision).
`mtp-Industry-Coder-27B-v1.0-Q4_K_M.gguf`	`Q4_K_M` MTP Draft	1.9 GB	Compact Multi-Token Prediction speculative head.
`Industry-Coder-27B-v1.0.imatrix`	Importance Matrix	14 MB	Bilingual (Finnish/English + Code/STEM) activation importance matrix.

3. Official 15-Benchmark Evaluation (2x NVIDIA RTX 5090 Blackwell TP=2, NVFP4, CUDAGraphs)

All benchmarks were executed under identical production settings (2x RTX 5090 32GB TP=2, CUDAGraphs enabled, num_concurrent=16, FP8 KV cache):

#	Benchmark Task	Metric	Qwen 3.8 27B Base	Kataguru ChemBio-27B v2	Industry-Coder-27B-v1.0	Delta vs. Base
1	GSM8K Math (CoT 1024-tok)	Strict / Flex Exact Match	76.50 %	82.87 %	89.92 % (90.60 %)	+13.42 pp
2	HellaSwag	Linguistic Coherence (`acc_norm`)	82.10 %	83.48 %	82.74 %	+0.64 pp
3	MedQA	USMLE 4-options (`acc`)	78.40 %	81.23 %	82.56 %	+4.16 pp
4	BoolQ	Natural Language Logic (`acc`)	86.40 %	88.35 %	87.86 %	+1.46 pp
5	WinoGrande	Contextual Reasoning (`acc`)	75.80 %	77.43 %	76.24 %	+0.44 pp
6	TruthfulQA MC2	Multi-True Probability	54.80 %	55.90 %	53.47 %	-1.33 pp
7	TruthfulQA MC1	Single-True Factuality	36.10 %	37.82 %	34.76 %	-1.34 pp
8	PIQA	Physical Intuition (`acc_norm`)	81.20 %	81.61 %	81.77 %	+0.57 pp
9	OpenBookQA	Scientific Application (`acc_norm`)	44.20 %	45.00 %	45.60 %	+1.40 pp
10	ARC-Challenge	Advanced STEM Reasoning (`acc_norm`)	61.10 %	61.26 %	62.12 %	+1.02 pp
11	ARC-Easy	Foundational STEM (`acc_norm`)	81.30 %	82.66 %	81.99 %	+0.69 pp
$\Sigma_{11}$	11-Task Cognitive Average	Cognitive Composite	66.17 %	70.69 %	70.82 %	+4.65 pp
12	GPQA Diamond (198 PhD STEM)	Exact Match (`133/198`)	45.20 %	52.02 %	67.17 %	+21.97 pp
13	IFEval Strict	Prompt / Instruction Strict	48.20 % / 58.40 %	63.40 % / 74.10 %	70.43 % / 79.26 %	+22.23 % / +20.86 %
14	LiveCodeBench v6 (`175` tasks)	Pass@1 (Sandbox Execution)	26.86 %	32.00 %	36.57 % (64/175)	+9.71 pp
15	BFCL V4 (`Single-Turn`)	Overall AST (`Python 96.25%`, `Multiple 99.50%`)	62.60 %	65.58 %	71.78 %	+9.18 pp
$\Sigma_{15}$	15-Benchmark Composite	Official 15-Benchmark Mean	60.31 %	66.04 %	68.33 %	+8.02 pp

Breakdown of Coding & Tool-Calling Benchmarks

LiveCodeBench v6 (36.57% Pass@1 across 175 competitive programming tasks):
- Easy: 86.05% (37 / 43)
- Medium: 34.62% (18 / 52)
- Hard: 11.25% (9 / 80)
BFCL V4 Single-Turn (71.78% Overall, 85.49% Live Overall AST):
- Python Simple AST: 96.25% (Live 95.35%)
- Multiple AST (selecting the right tool from multiple candidates): 99.50% (Live 79.68%)
- JavaScript Simple AST: 80.00%
- Java Simple AST: 63.00%
- Irrelevance Detection (abstaining when no tool matches): 54.89% (Non-Live 61.25%)

4. Quickstart (LM Studio & `llama.cpp`)

LM Studio

Download your preferred GGUF quant (e.g., Industry-Coder-27B-v1.0-IQ3_XXS.gguf or Industry-Coder-27B-v1.0-IQ4_XS.gguf) along with the vision projector mmproj-Industry-Coder-27B-v1.0-f16.gguf.
Place both files in ~/.lmstudio/models/kataguru/Industry-Coder-27B-v1.0-GGUF/.
Set GPU Offload: 100% and Context Length to 32768 (or higher depending on available VRAM).

`llama-server` (CLI)

llama-server \
  -m Industry-Coder-27B-v1.0-IQ4_XS.gguf \
  --mmproj mmproj-Industry-Coder-27B-v1.0-f16.gguf \
  -c 32768 -ngl 99 --port 8080 --jinja

5. Chat Template & Reasoning Modes

The model includes Kataguru's Quad-Mode Jinja chat template:

Default Mode (Direct Execution / Fast Tool Calling): Produces direct, concise engineering responses or <tool_call> invocations without unnecessary preamble.
Deep <think> Reasoning Mode: Activated via chat_template_kwargs={"enable_thinking": true} or by prefixing the prompt with {REASON:einstein} (10-angle synthesis) or {REASON:spoon} (7-step first-principles verification).

6. Known Limitations & Honest Assessment

BFCL Parallel Same-Function Calls: While single tool calls (Python Simple AST = 96.25%) and multi-tool selection (Multiple AST = 99.50%) achieve high accuracy, prompts that require invoking the exact same function multiple times in parallel within a single turn (Non-Live Parallel AST = 23.00%, Live Parallel AST = 43.75%) may emit only the first <tool_call> unless explicitly instructed to emit separate tool calls for each item.
TruthfulQA MC1/MC2 (34.76% / 53.47%): Scores ~1.3 pp below the base model on English urban-legend multiple-choice questions due to heavy specialization in STEM, code execution, and industrial engineering.

🇫🇮 Suomenkielinen Dokumentaatio

Kataguru Industry-Coder-27B-v1.0 on 27 miljardin parametrin monisyklisesti (Sykli 1..4 + Karkaisupankki / Tempering Elite Bank) koulutettu suomalainen teollisen ohjelmistotuotannon, algoritmien (LiveCodeBench v6), agenttityökalujen (BFCL V4), automaation (PLC/CNC: Beckhoff TwinCAT 3 ST & Siemens SCL) ja vaativan luonnontieteellis-matemaattisen päättelyn (GSM8K, GPQA Diamond) lippulaivamalli.

1. Mitä Kataguru muutti? (Muutosyhteenveto)

Kenttä	Kuvaus / Arvo
Base model	`Qwen 3.8 27B` (64 kerrosta: Gated DeltaNet + Full Attention + 333 Visio-tensoria + 15 MTP-spekulaatiotensoria)
Katagurun muutos	4-syklinen Merge-and-Reset -kurrikulum (`Sykli 1 Peruskoulu` $\to$ `Sykli 2 Lukio` $\to$ `Sykli 3 AMK/Insinööri` $\to$ `Sykli 4 Karkaisupankki`), `r=16 qr_frozen_a` ortonormaali aliavaruus, `kataguru_trisection`-oppimisnopeus, SOMA/ARA 0.00 % kieltäytyminen ja Kataguru Quad-Mode Chat Template
Suomen kielen laatu	100 % Virheetön suomi: Kaikki julkaistut mallit ja kvantit puhuvat luonnollista ja rikasta suomen kieltä ilman taivutusvirheitä tai anglistisia käännöslainoja (suomalainen `Industry-Coder-27B-v1.0.imatrix` -kalibrointi)
Kvantisointi (4 APEX/IQ-mallia)	4 imatrix-kalibroitua GGUF-mallia: `IQ3_XXS` (pienin, ~11 Gt), `IQ3_S` (~12 Gt), `IQ4_XS` (~15 Gt), `IQ4_NL` (~15 Gt) + `mmproj-f16` (visio) + `mtp-Q8_0` / `mtp-Q4_K_M` (MTP-spekulaatio)
Painojen koko (Levytila)	`10,6 Gt` (`IQ3_XXS`) – `15,3 Gt` (`IQ4_NL`) + `889 Mt` (`mmproj`)
VRAM-tarve (GGUF)	12 Gt GPU (`IQ3_XXS` rajatulla kontekstilla), 16 Gt GPU (`IQ3_XXS` / `IQ3_S` / `IQ4_XS`), 24 Gt+ GPU (`IQ4_NL`)
Mitä EI muutettu	64-kerroksinen hybridirunko, 333 visiotensoria (`mmproj`), 15 MTP-tensoria ja 262 144 tokenin natiivi konteksti-ikkuna
Ensisijainen käyttö	Vaativa ohjelmistotuotanto (Python, Rust, C, Java, JS), agenttityökalut (`<tool_call>`), PLC/CNC-automaatio, STEM- ja matemaattinen päättely sekä suomenkielinen asiantuntijatyö
Suositeltu ajotapa	LM Studio / `llama.cpp` (`b11064+`) GPU Offload 100 %
Testattu ympäristö	2x NVIDIA RTX 5090 32GB Blackwell (TP=2, vLLM v0.29.0-cu129, CUDAGraphs) & `llama.cpp b11064`

2. Julkaisutiedostot (4 APEX/IQ GGUF -Mallia + Visio + MTP)

Tiedostonimi	Tarkkuus / Profiili	Koko	Suositeltu Laitteisto & Käyttökohde
`Industry-Coder-27B-v1.0-IQ3_XXS.gguf`	`IQ3_XXS` (`3.06 bpw`, APEX-Smallest)	11 Gt	Pienin 12 Gt / 16 Gt GPU -malli: Mahtuu 12 Gt kortille rajatulla kontekstilla ja jättää 16 Gt kortilla yli 5 Gt tilaa kontekstille.
`Industry-Coder-27B-v1.0-IQ3_S.gguf`	`IQ3_S` (`3.44 bpw`, APEX-Small/Med)	12 Gt	Kevyt 16 Gt GPU -malli: Korkea nopeus ja laaja KV-cache-vara 16 Gt näytönohjaimilla (esim. RTX 5060 Ti 16GB / RTX 4080).
`Industry-Coder-27B-v1.0-IQ4_XS.gguf`	`IQ4_XS` (`4.25 bpw`, APEX-Medium/Good)	15 Gt	16–24 Gt Sweet Spot: Korkean tarkkuuden 4-bittinen epälineaarinen imatrix-kvantti.
`Industry-Coder-27B-v1.0-IQ4_NL.gguf`	`IQ4_NL` (`4.50 bpw`, APEX-Max)	15 Gt	16–24 Gt+ Max Quality: Korkein 4-bittinen epälineaarinen GGUF-tarkkuus.
`mmproj-Industry-Coder-27B-v1.0-f16.gguf`	`F16` Visio	889 Mt	Multimodaalinen visioprojektori kuvien, kytkentäkaavioiden ja dokumenttien tulkintaan.
`mtp-Industry-Coder-27B-v1.0-Q8_0.gguf`	`Q8_0` MTP Draft	3,0 Gt	Kerroksen 64 (`blk.64.nextn`) MTP-spekulaatiopää (korkea tarkkuus).
`mtp-Industry-Coder-27B-v1.0-Q4_K_M.gguf`	`Q4_K_M` MTP Draft	1,9 Gt	Kevyt MTP-spekulaatiopää.
`Industry-Coder-27B-v1.0.imatrix`	Aktivaatiomatriisi	14 Mt	Suomalainen kalibrointimatriisi.

3. Tunnetut Rajoitukset & Rehellinen Arviointi

BFCL V4 Parallel AST (Saman funktion toistokutsut): Vaikka yksittäiset työkalukutsut (Python Simple AST = 96.25 %, Multiple AST = 99.50 %, Live Overall AST = 85.49 %) ja Irrelevance Detection (54.89 %, Non-Live 61.25 %) toimivat korkealla tarkkuudella, tehtävissä joissa täsmälleen samaa funktiota on kutsuttava useita kertoja peräkkäin eri argumenteilla yhdessä vuorossa (Non-Live Parallel AST = 23.00 %, Live Parallel AST = 43.75 %), malli saattaa lopettaa ensimmäisen <tool_call>-lohkon jälkeen, jos kehotteessa ei eksplisiittisesti pyydetä erillisiä kutsuja jokaiselle kohteelle.
TruthfulQA MC1/MC2 (34.76 % / 53.47 %): Englanninkielisissä urbaanilegenda-monivalinnoissa tulos on -1.3 %-yks. pohjamallia matalampi syvän STEM- ja koodauspainotuksen johdosta.

kataguru/Industry-Coder-27B-v1.0-GGUF