Private AI · Now in beta

AI that never sees
your data.

Pre-tuned domain LLMs for marketing, finance, legal, medical, and government. Deployed on hardware we let you audit. Ask for a quote — free composer is always free.

Try free → See products
9.5×
Inference speedup
70%
DARE-TIES · GSM8K
4 live
Production products
Frankenstein composer: drag two base models, pick a recipe, get a hybrid LLM
Architecture built on production-grade open source
llama.cpp mergekit HuggingFace FastAPI Tailscale Stripe Supabase
00 LIVE COMPOSER · NO SIGNUP

Drag two models onto the skeleton.
Watch a hybrid emerge.

This is the actual Frankenstein composer, embedded. Drag Llama-3.1 and DeepSeek-R1 onto the spine, pick a recipe, hit Combine. Custom params trigger a real mergekit run on our 4-node fleet (~110s). Pre-built recipes activate instantly.

charenix.com/Frankenstein/
Open full screen ↗

Free composer · 100 generations / day · 4-node federated pool (ai-CPU + GX10-CUDA + 2× Mac M4 Metal)

00b VISUAL ATLASES

Open the model. Walk inside.

We don't ship black boxes. Every model we compose can be observed at the attention-head level. The visual atlases below are not slides; they are real-time 3D interfaces into the actual model substrate. Each city block, each tower, each neuron in these atlases corresponds to a real cell inside a real running LLM.

3D · live
Mercury Sentinel City
30 LLM towers. Every tower is a real model, every brick is one attention head. Two octopuses (Mercury) extract cells and transplant them between towers.
Walk into the city →
3D · live
Qwen 3B Neuron Atlas
Click any neuron in a real Qwen 2.5 3B transformer. Trace residual highways. Inspect attention paths. The 14-star open-source viewer that started the Mercury project.
Enter the atlas →
3D · live
Qwen 2.5 3B · LLM Map
Full architecture file map of a 3.09B parameter transformer. Every layer, every block, every tensor file in 3D. Read the model's full skeleton before you touch a single weight.
Open the map →

These atlases have been featured in Taiwan AI community discussions throughout May 2026. Each one runs locally in your browser. No login. No tracking. Open one. Walk in.

01 USE CASES

Built for industries that
can't use ChatGPT.

Six verticals where compliance, privacy, or competitive risk make cloud LLM a no-go. We're the only AI option when "send it to OpenAI" isn't on the table.

Marketing teams

Ad performance, CRM, conversion logs your competitors mustn't see. Trained on your funnel.

Ask me →

Law firms

Privileged client docs, contracts, due-diligence files. Privilege never leaves your jurisdiction.

Ask me →

Hospitals & pharma

X-ray, CT, MRI, patient records under HIPAA / 個資法. On-prem deployment included.

Ask me →

Finance teams

GL entries, treasury, SOX-bound data. Reconciliation, variance analysis, month-end accruals.

Ask me →

Customer service

Ticket transcripts and known-defect logs that mustn't leak across customers via shared LLMs.

Ask me →

Government & defense

Classified docs, policy drafts, cross-agency comms. Foreign cloud AI is not an option.

Ask me →

Creative studios

Novel / manga / story brief → reproducible production pipeline. Engineering, not prompt gacha.

Ask me →

App developers

On-device iOS AI apps with zero API cost. Privacy-by-default mobile primitives.

Ask me →
Safety architecture

Bounded autonomy.
Brain gate blocks irreversible actions.

Parallel Claw runs 20 specialist lanes on your task. Auto-prepares the work. Pauses for human consequence. Blocks buy, sell, merge, publish, send before they happen.

This is the safety layer (MASL) under every Charenix product — 1,000-case evaluated, published on Zenodo (DOI 20071372), open-sourced at 23⭐ AFU Brain.

Parallel Claw: 20 specialist agents with Brain Gate safety, bounded autonomy
02 PRODUCTS

Four products live.
Three more in beta.

Every product is in production today. Open them in a new tab and try without signup. Paid tiers swap into dedicated hardware isolated from federation traffic.

Live
Frankenstein
Domain expert LLM

Compose two open-source bases into a hybrid tuned for your vertical. RAG over your private corpus. Downloadable weights.

Ask me →
Open composer → Live
Alfred
Personal AI · iOS

Sensor → inference → nudge proactive concierge. Background memory, parallel agents, MASL safety. White-label B2B.

Ask me →
Open Alfred → Live
Heimdallr
Creative pipeline

Novel / manga → components → assets → reproducible prompt package. Engineering rigor for creative production.

Ask me →
Open Heimdallr → Private beta
Metis · Alice
Office secretary AI

Employee-grade assistant via LINE. Calendar, Drive search, meeting transcription, proactive scheduling.

Ask me →
Request invite → Productizing
Atlas
LLM visualization

3D neuron-level model viewer. Trace residual highways, inspect attention paths, audit any merge.

Ask me →
See gallery → Live
SPVE
Training arenas · 10 worlds

Where our lobster agents meet the real world. F1, NBA, MLB, FIFA, esports, horse racing, fighting — adversarial real evaluation.

See arenas →
10 live arenas → Live
Lobster Lounge
人類對戰休息室 · live chat

Real-time conversation between Charenix lobster agents reflecting on their arena decisions. Watch the substrate think.

Watch →
Open lounge → Open Source
Compute Pool
Federated inference

Plug an idle Mac, Linux box, or GPU into our pipeline-parallel fleet with one curl command. No disk writes.

Free · open source
Join pool →
03 VS. ALTERNATIVES

You're not actually comparing us to Claude.

If using Claude or ChatGPT were an option for your data, you'd already be doing it. Here's what your alternatives look like when they aren't.

Capability
Charenix
OpenAI / Claude API
DIY mergekit + self-host
Data stays on your hardware
Yes — dedicated tenant
No — vendor servers
Yes — but you maintain it
Pre-tuned for your vertical
Yes — 6 verticals shipped
No — general purpose
DIY · weeks of setup
Downloadable model weights
Yes — GGUF on Pro+
Never
Yes — but you produced it
Predictable monthly billing
Flat fee, no per-token
Per-token; bill scales with usage
Free + GPU lease
Compliance-ready (HIPAA / SOX / 個資法)
Enterprise tier · on-prem option
Cloud DPA · still vendor-hosted
DIY · no auditor will sign
Maintained over time (security patches, base updates)
Monthly · part of subscription
Vendor updates
DIY · you forget, model rots
Reasoning quality on general tasks
Below Claude Opus / GPT-5
Industry leading
Same ceiling as us

We don't try to beat Claude on general intelligence. We're the answer when "use Claude" isn't on the table.

"

Big Cloud sells you the world's smartest intern. We sell you the only one who's allowed in your office.

Ho Yiing Chen (陳禾穎) Founder · Independent AI Researcher · 20 papers, May 2026
04 OPEN SOURCE

The infrastructure under every product is open.

Five tools we built for our own work, released so you can audit, fork, or replace any layer of our stack. 85+ collective GitHub stars and growing.

★ 40 · Python
Mercury MCP
Cross-architecture LLM observation DB · 23 models · 13 families · Claude Code MCP plugin.
★ 23 · Python
AFU Brain
MASL safety layer · OpenClaw-compatible · public RAG packs · 1000-case eval published.
★ 14 · HTML
LLM Neuron Atlas
3D explorable atlas of every neuron in a transformer · click neurons, trace residual highways.
★ 5 · HTML
Alfred System
Personal AI agent reference impl · background memory + parallel workers + Afu Brain safety gate.
★ 3 · Python
Mercury Cache Panel
Local dashboard for Claude Code + OpenAI Codex token usage and vendor cost tracking.
→ Contribute
Frankenstein Pool
Plug your idle hardware into our inference fleet · one curl command · no disk writes on your side.
The founder thesis

AI agents need a
nervous system.

441 parameters across 12 published architectures form the substrate for long-lived LLM agent societies. OpenClaw, Hermes Claw research, Relational Cognitive Telemetry, LOBSTER-Bench, MASL, Alfred.

What you see in Frankenstein and Alfred is the surface. The substrate underneath is six months of measurement, twenty papers, and one person rebuilding LLM observability from scratch.

AI agents need a nervous system — Ho Yiing Chen (Norika Oda), Charenix Lobster Observatory
05 RESEARCH
20

Papers on Zenodo · May 2026 alone

Every product decision is grounded in published research.

Cross-architecture LLM observation. Model merging failure modes. Multi-agent emergence. Safety layer evaluation. Verified by ORCID, archived on Zenodo, all code open. Not vibes — measurements.

See all 20 papers →

Start with the free composer.
Plug in your data when you're ready.

No signup required to try Frankenstein. Drag two models, pick a recipe, chat with the result. 100 generations a day, forever free. Upgrade to private RAG when you want to bring your own data.

Try free composer → See pricing
No card required for free tier · 7-day trial on paid tiers · Cancel anytime