Kataguru-Sceptic-35B-A3B-Basque-v3.2-NVFP4 — Blackwell Native NVFP4 (20.7 GiB / Single 24 GB GPU Ready)

🏢 Official Release by Datasein Oy (Finland) Developed and published by Datasein Oy (@kataguru, Chief Architect: Jari Katajisto, Finland).

Kataguru-Sceptic-35B-A3B-Basque-v3.2 is a sovereign Basque (Euskara) Industrial Coder (PLC/CNC: 96.67 %), Systems Software Engineer & 6-in-1 Multi-Mode / Multimodal Mixture-of-Experts (254 + 2 = 256 experts per layer) flagship model built on the Kataguru Sceptic 35B-A3B v3.1 foundation (15-Benchmark Mean: 70.26 %, GPQA Diamond: 70.20 %, LiveCodeBench v6: 39.43 %, BFCL V4 AST: 83.67 %, Industrial PLC/CNC: 96.67 %).

⚙️ Not Just a Language Model — A Full-Stack Industrial & Systems Coder in Basque (96.67 % PLC/CNC) Because 254 frozen Sceptic core experts (#0..#253) co-activate in every Top-8 token step alongside the 2 native Basque experts (#254, #255), Kataguru-Sceptic-35B-A3B-Basque-v3.2 is a production-grade Industrial Automation & Systems Software Coder that writes, audits, and debugs PLC/CNC and systems code directly in Basque (Euskara), Finnish, or English.

🛠️ Specialized Industrial Coding & Engineering Skills (#0..#253 + #EU)

  1. IEC 61131-3 PLC Programming (96.67 % Verified Benchmark):
    • Beckhoff TwinCAT 3 Structured Text (ST): Deterministic state machines (CASE), motion control (MC_MoveAbsolute, MC_Power, MC_GearIn), EtherCAT distributed clocks, and ADS diagnostics.
    • Siemens TIA Portal (SCL / S7-1500): Function Blocks (FB/FC), instance Data Blocks (DB), PID closed-loop control (PID_Compact), and PROFINET IO diagnostics.
  2. CNC G-Code, Post-Processors & Coordinate Kinematics:
    • Fanuc / Siemens Sinumerik 840D / Heidenhain: Multi-axis lathe & milling G-code (G41/G42 tool radius compensation, G54–G59 work offsets, canned cycles G71/G76/G81–G89), spindle/feedrate load verification, and collision-free toolpath auditing.
  3. Industrial Protocols, Fieldbuses & SCADA/MES Diagnostics:
    • OPC UA (IEC 62541), Modbus TCP/RTU, EtherCAT, PROFINET, CANopen, and SQL Server / WCF production telemetry troubleshooting.
  4. Functional Safety & Machinery Standards:
    • ISO 13849-1 (PLa–PLe), IEC 62061 / IEC 61508 (SIL 1–SIL 3), Safe Torque Off (STO), two-channel emergency-stop interlocks, and watchdogs.
  5. Systems, Embedded & Algorithmic Software Engineering (LiveCodeBench v6: 39.43 %, Easy: 90.70 %):
    • Modern C++20, Rust, C (embedded/RTOS), Python 3.12, C#/.NET, SQL, and Bash/Linux kernel/network hardening, concurrency, and memory-safe systems architecture.
  6. Deterministic <tool_call> AST Agent & Schematic Vision (BFCL V4: 83.67 %):
    • Reads P&ID diagrams, electrical wiring schematics, HMI error screens, and engineering plots via its 27-layer ViT Vision Encoder, and executes multi-step <tool_call> AST tool chains (0 syntax errors).

Unlike single-purpose language fine-tunes, Kataguru-Sceptic-35B-A3B-Basque-v3.2 preserves 100 % of Kataguru Sceptic's native multi-mode operating system, 27-layer multimodal vision encoder, <tool_call> AST agentic engine, and graduate STEM/industrial coding core inside a single model — switching roles dynamically per request with zero server restarts:

Official Basque (v3.2) Release Suite

  1. Full BF16 Master Weights (~67.5 GB): kataguru/Kataguru-Sceptic-35B-A3B-Basque-v3.2
  2. Blackwell Native NVFP4 (~20.9 GiB, Single 24 GB GPU Ready): kataguru/Kataguru-Sceptic-35B-A3B-Basque-v3.2-NVFP4
  3. Universal vLLM W4A16-AWQ (~19.5 GiB, Single 24 GB GPU Ready): kataguru/Kataguru-Sceptic-35B-A3B-Basque-v3.2-W4A16-AWQ
  4. GGUF APEX-Trio (APEX-Small, APEX-Medium, APEX-Good + mmproj — 16 GB GPU Ready): kataguru/Kataguru-Sceptic-35B-A3B-Basque-v3.2-GGUF

1. What Kataguru Changed (254 + 2 Sovereign Language MoE Architecture)

Field Description / Value
Base model kataguru/Kataguru-Sceptic-35B-A3B-v3.1 (254 frozen core experts #0..#253, 3B active parameters per token)
Kataguru v3.2 architecture 254 + 2 = 256 Experts per MoE layer (40 layers):
• Experts #0..#253 (100 % Frozen Core): Preserves the full scientific reasoning (GPQA Diamond 70.20 %), competitive coding (LiveCodeBench v6 39.43 %), agentic tool calling (BFCL V4 83.67 %), and industrial PLC/CNC capabilities (96.67 %) with zero catastrophic forgetting.
• Expert #254 (Basque Grammar, Morphology & Technical Specialist): Ergative-absolutive case alignment, polypersonal auxiliary verb agreement (Nor-Nori-Nork), and agglutinative postpositional suffixes, plus industrial engineering, PLC/CNC, clinical medicine, and statutory/patent terminology.
• Expert #255 (Basque Literature, Nonfiction & Bidirectional Translator): Full-page literary prose, modern nonfiction, natural dialogue, and faithful EN <-> EU translation (/translate / /itzuli).
Language Bank Training Set 900 human-verified golden samples (100 % Pass on Mechanical & Quality Gate) + contrastive router calibration (--moe-mode new_experts_only --recalibrate-router, r=32, alpha=64)
Dedicated Chat Template & Calibration Native Basque Multi-Mode chat_template.jinja (/eu, /translate, /itzuli, /erraza, [INSPECT], [CURATE], [SURGEON], [DECISION]) + Basque quantization calibration (calib_nvfp4_awq_256_eu.jsonl & imatrix_calibration_eu.txt)
Format & Footprint Blackwell NVFP4 (compressed-tensors, ~20.7 GiB on disk, ~19.4 GiB in vLLM --language-model-only — fits on a single 24 GB GPU!)
What was NOT changed All 254 core reasoning/STEM/coding experts (#0..#253), 27-layer multimodal vision encoder (333 tensors), Layer 40 MTP speculative head (785 tensors), 262,144 context window

2. Native Sceptic Multi-Mode, Multimodal & Agentic Capabilities (6 Engines in 1 Model)

Every Kataguru-Sceptic-35B-A3B-Basque-v3.2 checkpoint includes the complete Kataguru Sovereign Multi-Mode Master Architecture out of the box:

Built-In Mode / Engine Trigger Tag or Template Arg Native Capability & Production Role
1. Forensic Quality Inspector [INSPECT] or sceptic_mode="inspect" Audits code, math derivations, and SFT datasets; emits deterministic verdict-first JSON ({"verdict": "ACCEPT|REJECT|REVIEW", "fatal_error": false, "learning_value": 9.2, "score": 9.2, ...}) with <100 ms early-exit rejection.
2. First-Principles Data Curator & Surgeon [CURATE] / [SURGEON] Synthesizes high-learning-value (>= 8.5) training samples from raw technical documents ([CURATE]) or self-heals buggy/truncated code and flawed proofs ([SURGEON]).
3. 20 ms System 1 Reflex Judge [DECISION] or sceptic_mode="decision" Closes <think> automatically and scores candidates via single-token (max_tokens=1, logprobs=True) 0–9 / A/B logprob distributions (>1M samples/day throughput).
4. Full-Page Bidirectional Translator /translate or /itzuli Zero-chatter, paragraph-preserving literary, legal, medical, and engineering translation (EN <-> EU & FI <-> EU), keeping LaTeX formulas and code blocks intact (0.00 % refusal).
5. Native Basque & Plain-Language Control /eu, /en, /fi, /erraza Locks response language across all modes (including code comments and [INSPECT] rationales), or switches to accessible plain Basque (/erraza).
6. Multimodal Vision, <tool_call> Agent & PLC/CNC Default Chat + images / tools=[...] 27-Layer ViT Vision (schematics, charts, UI screenshots, OCR, video), BFCL V4 83.67 % <tool_call> AST, Systems SWE (C, C++20, Rust, Python), and Industrial Automation (96.67 %) (Beckhoff TwinCAT 3 ST, Siemens TIA Portal SCL, CNC G-code, OPC UA).

Empirical Cognitive Reasoning Modes (chat_template.jinja)

Control reasoning structure and latency per prompt via inline tags (automatically stripped before rendering) or chat_template_kwargs={"reasoning_effort": "..."}:

  • {REASON:ihyper} (Instant MECE): +8.69 pp higher accuracy and -31 % faster (0 <think> overhead; structured MECE synthesis directly in the answer).
  • {REASON:hyper} / {REASON:logic}: Deliberate MECE decomposition or formal first-principles axiom/unit verification inside <think>...</think>.
  • {REASON:007} (Fault-Tolerant Agentic Loop): Deterministic state-machine execution for <tool_call> AST tool use and PLC/CNC control logic.
  • {REASON:spoon} / {REASON:einstein} / {REASON:help} / {REASON:off}: Deep premise validation, multi-angle scientific synthesis, interactive mode guide, or instant zero-think reply.

3. Verified Benchmark & Live Audit Summary

Benchmark / Capability Pillar Score Notes
15-Benchmark Core Foundation Mean 70.26 % Inherited 100 % intact via frozen Experts #0..#253
GPQA Diamond (PhD-level STEM) 70.20 % (139 / 198) Graduate physics, chemistry & molecular biology
Industrial PLC / CNC / TwinCAT / SCL 96.67 % IEC 61131-3 Structured Text, Siemens SCL & G-code verification
BFCL V4 Single-Turn AST (Tool Calling) 83.67 % (Non-Live: 88.44 %) Native <tool_call> AST & irrelevance detection (80.26 %)
LiveCodeBench v6 (Pass@1) 39.43 % (Easy: 90.70 %) Competitive algorithmic programming
Live Basque 10/10 Audit (v3.2) 10 / 10 PASS (100 %) HiTZ EvalEU Suite (EusProficiency, EusExams, EusReading, EusTrivia) and Full-Page Literary & Technical Translation (EN/ES <-> EU)

4. Honest Limitations & Hardware Sizing

  • High-Temperature Creative Fiction: For strict morphological and proofreading accuracy in Basque, use temperature=0.2–0.3 with presence_penalty=0.3.
  • VRAM Footprint:
    • W4A16-AWQ (~19.5 GiB disk / ~18.0 GiB VRAM) & NVFP4 (~20.9 GiB disk / ~19.4 GiB VRAM): Run on a single 24 GB GPU (RTX 3090 / 4090 / 5090 / L4) in vLLM (--language-model-only) or 2x 16–32 GB GPUs with full multimodal vision!
    • GGUF APEX-Small / Medium / Good (~11.9–15.4 GiB): Run on a single 16 GB GPU or Apple Silicon Mac (llama.cpp, LM Studio, Ollama).
    • Full BF16 (~67.5 GB): Requires ~72 GB VRAM (2x 48–80 GB).

Laburpena, Industria-Programazioa eta Sceptic Gaitasun Natiboak (Euskaraz)

Kataguru-Sceptic-35B-A3B-Basque-v3.2 ez da hizkuntza-eredu arrunt bat soilik, 256 adituko MoE arkitektura bakarrean integratutako Industria-Programatzailea (PLC/CNC: 96,67 %) eta 6-batean sistema multifuntzional subiranotasunezkoa baizik (254 Sceptic oinarrizko aditu izoztu #0..#253 + 2 euskara batuko aditu espezializatu #254 eta #255):

  • Industria- eta Sistema-Programatzaile Osoa:
    • PLC eta industria-automatizazioa (IEC 61131-3: 96,67 %): Beckhoff TwinCAT 3 Structured Text (ST), Siemens TIA Portal (SCL / S7-1500), egoera-makinak, higidura-kontrola, PID kontrolagailuak eta EtherCAT/PROFINET diagnostikoa.
    • CNC G-kodea eta mekanizazioa: Fanuc / Sinumerik 840D / Heidenhain (G41/G42 erradio-konpentsazioa, G54–G59).
    • Industria-protokoloak eta segurtasun funtzionala: OPC UA (IEC 62541), Modbus TCP/RTU, ISO 13849-1 (PLe), IEC 62061 (SIL 3 / STO).
    • Sistema- eta algoritmo-programazioa (LiveCodeBench v6: 39,43 %, Easy: 90,70 %): C++20, Rust, C, Python 3.12, C#/.NET, SQL eta Linux/Bash.
  • Euskara Batua eta orrialde osoko itzulpena (/eu, /translate, /itzuli, /erraza): Ergativitatea eta Nor-Nori-Nork aditz-komunztadura zehatza, terminologia teknikoa eta bi noranzkoko itzulpena.
  • Sceptic-en 6 lan-modu natiboak: [INSPECT], [CURATE] & [SURGEON], [DECISION] (<20 ms), 27 geruzako ViT ikusmen multimodala, <tool_call> agentea (BFCL V4: 83,67 %) eta {REASON:ihyper} / {REASON:007}.

Suomenkielinen Yhteenveto / Finnish Summary

Kataguru-Sceptic-35B-A3B-Basque-v3.2 on Kataguru Sceptic 35B-A3B (v3.1, 15-benchmark: 70,26 %) -pohjalle rakennettu suvereeni 254 + 2 = 256 ekspertin Teollisuuskoodari (PLC/CNC: 96,67 %) ja 6-in-1-monitoimimalli (Basque / Euskara):

  1. Täyden tehon Teollisuuskoodari & Järjestelmäohjelmoija (Ei pelkkä kielimalli):
    • PLC- ja automaatio-ohjelmointi (96,67 %): Kirjoittaa, tarkastaa ja korjaa Beckhoff TwinCAT 3 Structured Text (ST)- ja Siemens TIA Portal (SCL / S7-1500) -ohjelmia (tilakoneet, liikeohjaus, PID-säätö, EtherCAT/PROFINET-diagnostiikka).
    • CNC G-koodi ja koneistus: Fanuc / Sinumerik 840D / Heidenhain -työstöradat (G41/G42-teräsädekompensaatio, G54–G59, kiinteät työstökierrot ja törmäystarkastus).
    • Teollisuusväylät ja koneturvallisuus: OPC UA (IEC 62541), Modbus TCP/RTU, ISO 13849-1 (PLe) ja IEC 62061 (SIL 3 / STO).
    • Järjestelmä- ja algoritmikoodaus (LiveCodeBench v6: 39,43 %, Easy: 90,70 %): Tuotantotason C++20, Rust, C, Python 3.12, C#/.NET, SQL ja Linux/Bash suoraan kohdekielellä (Euskara), suomeksi tai englanniksi.
  2. Natiivi Sceptic-monitoimiarkkitehtuuri yhdessä mallissa: Sisältää jäädytetyn 254 ekspertin tieteellisen ja teknisen ytimen (GPQA Diamond 70,20 %, BFCL V4 <tool_call> 83,67 %), 27-kerroksisen multimodaalisen vision (mmproj mukana kytkentäkaavioiden ja HMI-ruutujen lukuun), laaduntarkastajan ([INSPECT]), aineistosyntetisoijan ja koodikirurgin ([CURATE] / [SURGEON]), <20 ms Systeemi 1 -pikatuomarin ([DECISION]) sekä empiiriset {REASON:ihyper} / {REASON:007} -päättelytilat.
  3. Kaksi dedikoitua kielieksperttiä (#254 ja #255, 900 näytettä):
    • Ekspertti #254: Kohdekielen (Euskara) taivutusoppi, morfologia sekä tekniikan, automaation, lääketieteen ja lakikielen erikoissanasto.
    • Ekspertti #255: Kohdekielen kaunokirjallisuus, tietokirjallisuus ja koko sivun kaksisuuntaiset käännökset (/translate / /itzuli).
  4. 24 GB vLLM & 16 GB GGUF -valmis: Täydet BF16-painot, yhdelle 24 GB näytönohjaimelle mahtuvat NVFP4 (~20,9 GiB) ja W4A16-AWQ (~19,5 GiB) sekä kielikohtaisella imatrix-korpuksella kalibroidut GGUF APEX-kvantit (APEX-Small, APEX-Medium, APEX-Good + mmproj-f16).
  5. Julkaisija (Datasein Oy, Suomi): Kehittäjä Datasein Oy (@kataguru, pääarkkitehti Jari Katajisto).
Downloads last month
120
Safetensors
Model size
36B params
Tensor type
BF16
·
U8
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for kataguru/Kataguru-Sceptic-35B-A3B-Basque-v3.2-NVFP4

Quantized
(40)
this model

Space using kataguru/Kataguru-Sceptic-35B-A3B-Basque-v3.2-NVFP4 1

Collection including kataguru/Kataguru-Sceptic-35B-A3B-Basque-v3.2-NVFP4