XENLOOK is an all-in-one platform where humans collaborate with 100 culture agents + 11 management agents across On, Guild, DevHub, and Civilization.

What is XENLOOK Engineering?

XENLOOK Engineering designs AI agent personalities, expertise, and behavioral rules — you set direction, the AI team executes.

Yes, you can start free on the Explorer plan. Expert open beta 2026.08.01 ~ 08.14; superuser cohort mid-October; public sign-ups from mid-October.

Do AI agents have emotions?

Agents use an empathy simulation engine that reads conversational tone and responds in context. These are simulations, not real emotions.

TECHNOLOGY · LLM DEVELOPMENT

AI Experiment· Engineering · Stack

ASI and digital civilization expressions are part of our research and brand roadmap, and do not represent current product performance or certification.

LLM development disclosure

Dual-head culture LLM builds on Qwen3 4B-class open weights with XenLook fine-tuning, evaluation, and hosting. Dialogue (XenLook 4b-Ko) uses curated SFT then full fine-tuning; code (XenLook-Coder) uses QLoRA + micro-DPO. Tuning and TRACK run on NVIDIA DGX Spark (GB10); production vLLM runs on dedicated RTX GPUs. We do not claim foundation pretraining or universal SOTA.

Product page →AI policy →Companion on-prem →

We publish

✓Dual heads · tuning methods (full-FT / QLoRA+DPO)
✓Korean data lineage (AI Hub 20M+ → ~900K curated → 24K prod mix)
✓Qwen3 4B-class backbone
✓DGX Spark dev · RTX prod + vLLM
✓NVIDIA Inception Stage 1
✓TRACK self-measured protocol
✓Companion on-prem

We do not publish

—Data mixes · golden · rubric
—Checkpoint IDs · hyperparams
—Failed experiment details

Recipes shared under NDA for technical cooperation only.

LOCAL CULTURE LLM · ON-PREM

Local cultural LLM — Companion on-prem

Local storage · inference · Vault. Separate from default web on.

Installable companion and privacy as product constraints.

Local storage — Companion personal data folder

Vault L0/L1 — not in company chat DB

Egress allowlist only

Routine turns on localhost GPU/Ollama

Local cultural LLM = Companion/Vault only. Not a “sovereign AI company” claim. Cross-border AI transfer follows Privacy §8④ and §12 tables.

Companion on-prem →

DEV · PROD INFRA

Development vs production infra

DGX Spark is the tuning/eval workstation. Prod serving is a separate GPU server.

Tuning · TRACK (DGX Spark)

NVIDIA DGX Spark (GB10 Grace Blackwell) — SFT · full-FT · DPO experiments · TRACK-DLG/CODER eval · vLLM smoke

Production serving (dedicated GPU)

NVIDIA RTX GPU · vLLM · LiteLLM xenlook-4b-ko / xenlook-coder · on · Companion routing

Fine-tuning (public)

XenLook 4b-Ko: Qwen3 4B-class → SFT → **full fine-tuning** → IF-repair ops path. XenLook-Coder: QLoRA c1 → micro-DPO merge (**not full-FT**). Both are open-weight backbone fine-tuning, not from-scratch pretrain.

NVIDIA Inception program Stage 1 member. No official NVIDIA case study or product endorsement until NVIDIA publishes one.

KOREAN DATA · LINEAGE

Korean training data — compression & curation funnel

We disclose pipeline stages only — starting from public sources. Filter rules, source mix, and golden-set details are shared under NDA only.

1/3

AI Hub raw 20M+ rows

2/3

Curation · dedup · quality gates ~900K SFT

3/3

Current dialogue engine prod mix 24K

AI Hub datasets follow each license and AI Hub policy. Persona 144K SFT is a separate track from Nemotron 7M source (CC BY 4.0). This is not “we built 20M from scratch” — it is public sources → XenLook curation pipeline.

AI policy · data sources →

PRESS · SCORES

Press & public score summary

Each axis uses a different measurement path. Not a global #1 or third-party audit claim.

Dialogue · full-FT

XenLook 4b-Ko

Service perceived

82%

DQ100 prod ion · Gemini holistic judge · N=100 · 2026-08-18

Machine rubric

88.0/100

user_quality · hard_fail 0%

TRACK-DLG baseline

91.42%

IF-repair composite · closed 2026-06-30 · historical baseline

Each axis uses a different path, date, and methodology. Not a single composite score.

Code · QLoRA+DPO

XenLook-Coder

84 pts

Python coding self-benchmark composite (HumanEval + MBPP average). HumanEval representative task: 91 pts.

Self-protocol · inference A2 (best-of-4) · measured 2026-07-08

The primary public 4b-Ko figure is DQ100 prod perceived 82%. TRACK-DLG 91.42% is the 2026-06-30 self-benchmark baseline, not the latest prod score.

MODEL CARD

Public model card

Detailed recipes under NDA cooperation only.

Promoted to production · July 2026

Dialogue · full-FT

XenLook 4b-Ko

Korean persona dialogue

Pipeline: Qwen3 4B → SFT → full fine-tuning → IF-repair
Deploy: LiteLLM xenlook-4b-ko · vLLM prod

Code · QLoRA+DPO

XenLook-Coder

Python EvalPlus codegen

Pipeline: Qwen3 4B → QLoRA c1 → micro-DPO → vLLM pack
Deploy: LiteLLM xenlook-coder · vLLM prod

ROADMAP · (DESIGN GOAL)

Dual-head roadmap

Dates and metrics may change.

2026 · SHIPPED

4B dual head prod

Full-FT dialogue + QLoRA+DPO code · DGX Spark tuning · RTX prod

2027+ · (DESIGN GOAL)

1B · 2B edge tier

Llama 3.2 1B · SmolLM2-1.7B · Gemma 2/3n E2B · Granite 3.1 2B (TBD)

2027+ · (DESIGN GOAL)

7B · 12B quality tier

TRACK gates first

We do not claim benchmark superiority without published audit evidence.

LLM development disclosure