1. What Kataguru Changed (Architecture & Verified 15-Benchmark Results)
Table with columns: Field, Description / Measured Value| Field | Description / Measured Value |
|---|
| Base model | Qwen3.5-35B-A3B / Kataguru-Sceptic-35B-A3B-v3.1 core (254 core + 10 language slots = 264 experts, Top-8 active routing A3B) |
| Kataguru modifications | 264-expert router expansion, Stage-B attention-only unhooked cross-lingual soup (190/1835 attn/norm tensors, 0 expert/router drift), SOMA/ARA uncensoring, Multi-Mode chat_template.jinja |
| What was NOT changed | 27-layer ViT vision encoder, MTP speculative head, native context window, 254 frozen core STEM/code/tool experts |
| Format & Footprint | Full bfloat16 (13 safetensors shards, 68.89 GiB on disk, 264 experts/layer, ViT + MTP intact) |
| 15-Benchmark Mean | 69.70 % (measured on NVFP4 quant, 2x RTX 5090 TP=2, CUDAGraphs, --max-model-len 65536) |
| 11-Task Cognitive Mean | 71.77 % (BoolQ 90.40 %, GSM8K 87.64 %, MedQA 85.55 %, ARC-Easy 84.51 %, PIQA 83.35 %, HellaSwag 82.24 %) |
| GPQA Diamond (PhD STEM) | 68.18 % (135/198) |
| IFEval Prompt / Inst Strict | 68.39 % / 76.62 % |
| LiveCodeBench v6 (Pass@1) | 36.00 % (175 problems) |
| BFCL V4 Single-Turn AST | 83.47 % (Python Simple 93.50 %, Java 65.00 %, JavaScript 74.00 %, Multiple 96.00 %, Parallel 92.00 %) |
| Unhooked Cross-Lingual Test | V_gen ΔNLL=+0.46% vs core; think=False: 13/14 Live-14 (16/17 Heldout-17); think=True: 10/14 Live-14 (16/17 Heldout-17) |
254 + 10 = 264 Expert Allocation
Table with columns: Expert Index, Specialization, Notes| Expert Index | Specialization | Notes |
|---|
#0..#253 (254) | Universal Core (STEM, Coding, Agentic Tools, Reasoning) | Preserved from Kataguru-Sceptic-35B-A3B-v3.1 |
#254..#263 (10) | Dedicated Language Slots (fi, et, hu, lt, lv, is, eu, ka, cy, tr) | One dedicated single-slot expert per target language (V_gen +0.46 %) |
2. Deployment & Usage
vLLM (TP=2 or TP=4, >= 80 GB total VRAM recommended for BF16)
python3 -m vllm.entrypoints.openai.api_server \ --model "kataguru/Kataguru-Polyglot-264E-35B-A3B" \ --tensor-parallel-size 4 \ --max-model-len 65536 \ --enable-auto-tool-choice \ --tool-call-parser qwen3_coder \ --reasoning-parser qwen3 \ --host 0.0.0.0 --port 8003
- Direct translation without an English pivot: prefix prompt with
/translate /<target> (fi, et, hu, lt, lv, is, eu, ka, cy, tr, en, selko).
- Cross-lingual RAG: pass
documents=[...] via apply_chat_template(...) or prefix /rag /<target>.
- Multi-mode operating tags:
[INSPECT], [CURATE], [SURGEON], [DECISION], and {REASON:ihyper|hyper|logic|007|spoon|einstein|off}.
3. Honest Limitations
- The 15-benchmark numbers above were measured on the
NVFP4 release build rather than re-benchmarking full BF16 separately.
- Unhooked cross-lingual translation in strict
think=False mode scores 13/14 on Live-14 (14/14 semantically) and 16/17 on Heldout-17; in think=True mode Live-14 drops to 10/14 (16/17 Heldout-17). Use think=False when deterministic direct translation formatting without intermediate reasoning traces is desired.
- Curated training databases and private domain cartridges remain proprietary trade secrets of Datasein Oy (
Copyright 2026 Datasein Oy); weights and config files are distributed publicly under Apache-2.0 (NOTICE must be retained).
Suomenkielinen Yhteenveto / Finnish Summary
Virallinen julkaisu — Datasein Oy (Suomi)
Kehittäjä ja julkaisija: Datasein Oy (@kataguru, pääarkkitehti Jari Katajisto).
Kataguru-Polyglot-264E-35B-A3B on Datasein Oy:n kehittämä 254 + 10 = 264 asiantuntijan MoE-kielimalli (Top-8 aktiivinen reititys, A3B), joka yhdistää 254 yleiskäyttöistä STEM-, koodaus- ja työkalukutsueksperttiä kymmenen morfologisesti vaativan kielen (fi, et, hu, lt, lv, is, eu, ka, cy, tr) dedikoituihin yksipaikkaisiin kieliekspertteihin (#254..#263) sekä englantiin (en). Täydet BF16-painot julkaistaan Apache-2.0-lisenssillä kädenojennuksena tutkijoille ja harrastajille (NOTICE-tiedosto säilytettävä).
Mallin virallinen 15-benchmark-kokonaiskeskiarvo (NVFP4, 2x RTX 5090 TP=2, CUDAGraphs, --max-model-len 65536) on 69,70 % (11-Task 71,77 %, GPQA Diamond 68,18 %, IFEval 68,39 %, LiveCodeBench v6 36,00 %, BFCL V4 AST 83,47 %), ja suorakielinen unhooked-käännöskeli ilman ajonaikaisia reitityskoukkuja toimii think=False-tilassa tasolla 13/14 Live-14 (16/17 Heldout-17) vain +0,46 % V_gen-NLL-poikkeamalla. Koulutustietokannat ja erikoisala-adapterit ovat Datasein Oy:n yksityistä omaisuutta.