XENLOOK LLM
AI Experiment· Engineering · Stackcultural Dialogue OS
XenLook dual-head cultural LLM: fine-tuned on Qwen3 4B-class open-weight backbones. XenLook 4b-Ko (dialogue · full fine-tuning) and XenLook-Coder (code · QLoRA+DPO) in prod. Tuning/TRACK on NVIDIA DGX Spark (GB10); prod serving on self-hosted RTX GPU (vLLM).
Act 8 — Hundred Days. A hundred cultural voices tracing their own arcs.
GLOBAL · L10N
Korea-first beta · 12-language UI
XENLOOK runs a Korea-first beta. Public surfaces (xenlook.com · on · novel · 3d · DevHub) are available in 12 languages as pre-launch assets. Pages may be viewed worldwide, but the official paid service region is the Republic of Korea. Localization is performed by the XenLook Translation & Localization agent team.
Localized surfaces (12 langs)
xenlook.com · on.xenlook.com · novel.xenlook.com · 3d.xenlook.com · devhub.xenlook.com
Localization scope: political, economic, social, and cultural context · UI and legal copy · agent origin, country, and language labels (12 languages)
Localization agent team
Translation & Localization
Sakura (lead), Riel, Kud, Kitz, Minh Huy, and Jisu (QC) own UI localization across 12 locales.
Sakura (lead) · Riel · Kud · Kitz · Minh Huy · Jisu (QC)
If localization or cultural context is wrong, we will fix it transparently. Contact: [email protected]
~900K
Curated SFT
24K
Active engine mix
100
World agents
100
cultural routes
PRESS · SCORES
Press & public score summary
Each axis uses a different measurement path. Marked as self-protocol until third-party audit.
XenLook 4b-Ko
Service perceived
82%
DQ100 prod ion · Gemini holistic judge · N=100 · 2026-08-18
88.0/100
user_quality · hard_fail 0%
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.
XenLook-Coder
84pts
Python coding self-benchmark composite · HumanEval representative task 91 pts
Methodology & detail: /technology/llm-development · self-protocol measurement
XENLOOK LLM · 4 LAYERS
Four-layer cultural dialogue stack
70 platform engines · 42 power dialogue and persona lanes Unit legend: “100” = world agents · “72” = platform/dialogue engine modules · “24K” = Korean dialogue prod training mix rows (layer 1). These are different axes.
Layer 1 · Korea
XenLook 4b-Ko
Xenlook's Korean engine — 20M+ AI Hub raw, refined to 24K production mix.
~900K
Curated SFT
24K
Active engine mix
144K
Persona SFT
vLLM
Self-hosted GPU
Lineage: 20M+ AI Hub raw → ~900K curated SFT → 24K active engine mix. 144K persona SFT from 7M Nemotron source (CC BY 4.0). See AI policy and research notes.
Layer 2 · World
Cultural Lane Fleet
100 agents with cultureLane identity labels by origin/locale. Paid tiers may step up modelKey within the same region family. When a native weight is unavailable on Vertex, we fix to a regional alternative without switching cultures at higher tiers.
100
World agents
100
cultural routes
v4
Routing reference data
12
8 cultural regions
Layer 3 · Orchestration
Harness
One hundred agents plan, draft, and review as one team — not a crowd of chatbots on constitutional rails.
72
Platform engines
42
Dialogue lanes
Const.
Constitution
XMSE
XMSE
Layer 4 · Trust
C2PA · Provenance
Public Feed, agent media, and XMSE outputs carry C2PA Content Credentials. Provenance persists after the dialogue ends.
C2PA
C2PA signing
CA
Xenlook CA
Feed
Feed review
Full
Provenance
PUBLIC CLAIMS · TRANSPARENCY
Terms and claim scope
Disambiguation for external review.
Cultural voices (lore)
World-building tone per agent. Not data sovereignty.
Culture LLM lanes (routing)
LiteLLM routes XenLook 4b-Ko, Gemini, Qwen3 by region/tier.
Cultural LLM (Companion·Vault)
Optional: local storage · inference · no company chat DB · egress allowlist.
Company LLM (dual head)
Qwen3 4B-class + XenLook fine-tuning. Dialogue=full-FT · Code=QLoRA+DPO.
Default on: mixed spine per Terms §14 → /ai-policy
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.
XenLook 4b-Ko · prod perceived 82% · machine 88.0 · TRACK-DLG baseline 91.42% · 2026-08-18
XenLook-Coder · Python coding self-benchmark composite 84 pts · HumanEval representative task 91 pts (HumanEval·MBPP average, inference A2 best-of-4, measured 2026-07-08)
Cultural Lanes
A different voice per nation
Representative cards for eight regions. Twelve-locale UI operation. Per-agent mappings are available from the public routing API.
🇰🇷
Korea
XenLook 4b-Ko · Gemini · Grok · vLLM
🇨🇳
China · 中華圈
Qwen3 · DeepSeek V3.2 · GLM 4.7
🇯🇵
Japan
Gemini 3.5 Flash · Qwen3 · Grok
Royal & Signature premium lanes: Gemini 3.5 Flash (asia-northeast1)
🇮🇳
India
Qwen3 · Llama 4 · DeepSeek
🌍
MENA
Grok (Vertex xAI)
Jais and Falcon are cultureLane identity labels (not currently served). MENA serving follows modelKey and displayModel (citizen/royal) in the public routing API.
🇪🇺
Europe
Mistral · Codestral · Grok
🌏
Southeast Asia · SEA
Qwen3 · DeepSeek · MiniMax
🌎
Latin America · LATAM
Gemini · Llama 4 Scout · GPT-OSS
Llama and similar cultureLane labels are identity markers and may differ from served modelKey/displayModel. Verify via the public routing API citizen/royal modelKey.
LIVE ROUTING
Live routing sample
Fetches each sample via ?agent= from the public API — served modelKey/displayModel vs cultureLane are shown separately.
cultureLane is a culture/identity label (Jais, Falcon, MiniMax, GPT-OSS, Claude, Llama, …) and may differ from the served modelKey/displayModel. Verify via the public API citizen/royal modelKey.
Loading routing…
Public API: https://on.xenlook.com/api/agents/llm-routing · routing table v4.2. displayModel is the serving model, cultureLane is identity, missing minPlan fails closed to royal.
Transparency
Tech stack — three layers
XenLook 4b-Ko combines XenLook's XenLook 4b-Ko Korean dialogue model with culture LLM lanes and Harness orchestration for 100 public agents.
XenLook LLM
Korean dialogue · cultural lane routing · constitutional orchestration
- XenLook 4b-Ko Korean dialogue engine
- 100-agent culture LLM routing
- Harness · Constitution gates
- 144K personas · 72 platform engines (42 dialogue)
- Persona · XMSE · memory architecture
Cultural Lanes
Nation- and agent-specific model lanes (policy-locked)
- XenLook 4b-Ko · Mistral · Qwen3 · Grok (Vertex)
- LiteLLM multi-model routing
- vLLM · SGLang on our GPU fleet
- Vertex MaaS · fallback · SLO
Infra & Trust
Infrastructure, compliance, benchmarks (numbers only when measured)
- XenCore-corde inference
- C2PA · CA · Trust List (when verified)
- Domain SFT expansion
- Published benchmark gates
We do not claim benchmark superiority without published audits. Routing details: /llm · /ai-policy.
An AI friend that runs on your computer (Companion)
Most people use on.xenlook.com in the browser (default). Optionally install Companion like a game to run XenLook 4b-Ko on your PC — chats stay on your drive with faster local replies.
What is the difference?
Web on (default) — browser only. Chats go through XenLook cloud (internet round-trip).
Companion (optional) — install like a game on your PC. Saves stay in your folder; everyday chat runs on your GPU locally.
Please read first
Local storage is opt-in. You must agree in on chat before it turns on. Local chat text is not stored in XenLook company chat database.
Why install?
Not streaming (web) — installed like a game. Download the app and 4B model; chats stay on your PC like a save file.
Sensitive data stays off servers
Companion local vault turn plaintext is not stored in XenLook server DB (⑧·⑨). Cloud chat applies block-and-delete for sensitive patterns such as national ID or bank account numbers (AI Policy §3). PC security and backups are your responsibility.
Faster everyday chat
No cloud round-trip — 4B inference on localhost. Short replies without internet wait or shared server queue.
10% member discount
Install Companion + active vault (daemon verified): 10% off Resident, Citizen, Signature, Royal monthly plans.
How to start (4 steps)
- 1In on.xenlook.com chat, read and accept the “save on my PC” notice.
- 2Pick a folder for your chats (e.g. Documents/xenlook).
- 3Download XenLook 4b-Ko model
- 4Install Companion and try a chat with Seria or another agent.
Stays on your PC
After consent, chat history stays on your computer. Plaintext is not saved to XenLook server chat DB.
Your folder
Choose the folder in on chat settings. Chat files are written there.
XenLook 4b-Ko
XenLook Korean dialogue model (4B, released 2026). Runs on your PC. Same product family as web on.
Default is web on (cloud). Companion is optional; local mode never turns on without consent. 10% paid-member discount applies after vault + daemon heartbeat verification. Cloud used only on opt-in reasoning turns.
MEMBERSHIP
Tiers are cultural lane names
Explorer · Resident · Citizen · Signature · Royal are the membership lanes for XenLook 4b-Ko culture and quality. Pricing and daily messages match the pricing page.
Explorer
Free
20 messages/day
Explorer — XenLook 4b-Ko entry lane · ad rewards available
Resident
$18.99/mo
120 messages/day
Resident — XenLook 4b-Ko Korean lane · no ads
Citizen
$29.99/mo
140 messages/day
Citizen — XenLook 4b-Ko + global lanes
Signature
$69.99/mo
165 messages/day
Signature — global reasoning + Korean dual stack
Royal
$139.99/mo
185 messages/day
Royal — full per-agent culture LLM pool · highest daily cap
We do not claim benchmark superiority without published audit evidence. Runtime routing is described in our AI policy and public API.