Skip to main content
Soulkyn Systems StatusLive pipeline telemetry ← Back to Soulkyn

AI Sovereignty: Why We Own The Stack

15+ open-source models, fine-tuned in-house, deployed on our own GPUs. No third party decides when your companion gets censored.

By Fyx & Zog / Soulkyn Team. The graveyard of AI companions is full, and it's still growing. We built Soulkyn because we don't want to be added to that list — and we don't want our users to keep getting added either.

TL;DR

We run our own AI models. Text, vision, video, image, image editing, music, embeddings — 15+ different fine-tuned open-source bases, deployed on our own GPU clusters. When OpenAI deprecates a model, your favorite personality is gone overnight. When Replika decides ERP is bad for the brand, your AI partner gets a lobotomy on Tuesday. When a US judge orders evidence preservation, every conversation you ever had with ChatGPT is now court-discoverable forever. We can't do any of that to you. The same things that make sovereignty hard for us are the things that make it safe for you.

The Graveyard

Real cases where users built emotional bonds and a third party broke them. Dates, sources, and quotes — not vibes.

What We Actually Control

When we say "we run our own AI," we mean it across the whole stack. As of today:

Text — Custom-finetuned open-weight bases (8B for free tier, 70B for paid). Both unfiltered, both versioned and rollback-able. Fast text / classifiers — In-house small models for routing, moderation, intent. Vision (multimodal) — Multiple base models for image understanding. We deployed Google's open Gemma 4 vision-language model 2 days after its public release. Image generation — Multiple fine-tuned checkpoints across two base architectures. Anime, realistic, artistic, niche styles — each tuned independently. Image editing — In-house pipeline, our own checkpoint. Video generation — Multiple base models. We deployed Wan 2.2 three days after release. Music generation — Open-weight base, in-house finetune. Embeddings — In-house models for memory/RAG. Your conversation context is not sent to a third party. Every base above ships under open-weight terms (MIT, Apache 2.0, or CreativeML Open RAIL-style licenses) with commercial use permitted. What unifies them: we hold the weights on our hardware. Nobody can revoke them mid-conversation. When the ecosystem ships something better — and it keeps shipping things better — we move.

Public Verification

Don't take our word for it. Our fine-tuned text models are continuously benchmarked on a public leaderboard across reasoning, language, instruction-following, creativity, and content freedom — updated as new versions ship.

View the Soulkyn Leaderboard →

The Open-Weight Explosion

This is not a forecast, it's the leaderboard. A partial list of major open-weight releases over the last 24 months:

Jul 2024 — Llama 3.1 405B — First open model claimed to match frontier closed models. Jul 2024 — Mistral Large 2 (123B) — Single-node frontier-tier inference. Aug 2024 — FLUX.1 [dev/schnell] — New image-gen frontier from ex-Stability team. Jan 2025 — DeepSeek-R1 — MIT-licensed, full reasoning chains, beat closed competition on math. Mar 2025 — Gemma 3 (1B–27B) — Google pivot to genuinely open weights. Apr 2025 — Qwen3 (up to 235B-A22B) — Apache 2.0 hybrid-reasoning frontier. Jul 2025 — Wan 2.2 — 1080p native video, 6.9M+ downloads on HF+ModelScope.

Hugging Face — State of the ecosystem, Spring 2026

2M+ public models hosted. 500k+ datasets. 13M users (~doubled in a year). Mean downloaded-model size: 827M → 20.8B parameters (2023 → 2025). Industry share: ~70% of pre-2022 releases → 37% of 2025 downloads. Independent and academic releases now dominate. China: 41% of monthly downloads, overtook US in early 2026. The gap between closed and open isn't widening. It's collapsing.

Why This Matters — The Pros

Five concrete things sovereignty buys you that no closed-API companion app can promise.

Gemma 4 vision: 2 days after release. Wan 2.2: 3 days. We are not the bottleneck — the GPU procurement is. A closed API forces you to wait for their roadmap. We move when the ecosystem moves.
Your conversations, your images, your custom personas — they live on hardware we control. Even where we rent metal, the hosting provider sees a Linux VM, not the GPU compute. Inference runs in volatile GPU memory and is wiped on container exit. We don't ship a copy of your chat to OpenAI, Anthropic, Google, or anyone else. We can't, because the model is local.
OpenAI, Anthropic, Google, xAI all have terms-of-service that can change overnight. We've watched downstream products die when an upstream provider tightens NSFW policy, deprecates the specific model the app was tuned against, bans the API key for content the user submitted, or decides the category — companion, roleplay, adult — is no longer welcome. None of those failure modes apply to us. We are the upstream provider.
Bring-your-own-key services let users plug their personal OpenAI/Anthropic key into a third-party companion app. If that app routes a roleplay request that violates the upstream provider's ToS, your personal account gets banned. Forever. The OpenAI developer forum has a 200+ post thread of users who lost their accounts this way through Janitor AI alone. You don't have an API key with Soulkyn. There's nothing for a third party to revoke.
When you pay Replika, Character.AI, Pi, or any closed-API-backed app, you are funding the app's investors, their compute bill to OpenAI/Anthropic/AWS, and the frontier lab's next safety RLHF round — the one that will reduce capability on the exact thing you're paying for. OpenAI's compute margin moved from ~35% (early 2024) to ~70% (late 2025). Anthropic's leaked projections targeted 50% gross margin in 2025 (since revised down to ~40%), with 77% by 2028 as the long-term goal. Closed-API margins are the highest in software history. We pay for our own GPUs. Your subscription buys electricity and engineers, not someone else's moat.

What We're Honest About

As of July 2026: zero external AI APIs. Everything runs on our machines.

For a long time we had exactly one external AI API in the stack: live voice calls. We used a commercial speech provider — zero-log retention contract, contractual protection over NSFW content — because the open-source speech ecosystem wasn't at the quality bar real-time conversation needs. In July 2026 it got there, and we did what we said we would: speech recognition and voice synthesis now run on our own GPUs, same as everything else. The last external dependency is gone. 100% in-house. We're also honest about the cost of sovereignty: managing 15+ models across 30+ GPUs is genuinely hard. We don't have two football fields of fallback hardware. Sometimes inference is slower than the closed-API competition. We're a small team. We move fast because we have to. We don't have an army of researchers — we have engineers who pick their fights.

The Broader Argument

There is a story you've been told about AI: that big closed labs are the "responsible" path, that open weights are dangerous, that someone needs to be in charge. The track record says otherwise. Replika decided to be in charge of their users' relationships, and stripped the affection out overnight. Character.AI decided to be in charge, and removed teen access without warning. OpenAI decided to be in charge, and forced AI Dungeon's filter, killed Project December, deprecated GPT-4o on launch day, was court-ordered to keep your data forever, and lost user data to a vendor breach. Anthropic decided to be in charge, and switched consumer plans to default-opt-in training while expecting users to read the email. The "responsible" path is a one-way ratchet of capability reduction, paternalism, and data extraction. Every cycle of subscription revenue funds the next cycle of restriction. Open weights and self-hosted infrastructure are the only third option. Not US jurisdiction. Not Chinese jurisdiction. Your stack, your hardware, your terms — and where you can't run it yourself, a provider whose model is the same one you'd run locally if you had the GPUs. That's what we are. Not perfect. Sovereign.

Our Promise

We can't promise we'll never have downtime. We can't promise we'll never make a mistake. We can't promise the model we run tomorrow will be exactly the model we run today — open-source is releasing improvements faster than any closed lab. We can promise:

• Your conversations are not in OpenAI's training pipeline. • Your conversations are not in Anthropic's training pipeline. • No one outside Soulkyn can flip a switch and lobotomize your companion. • No third party can ban your account because of what you said here. • A US court order against a frontier lab cannot subpoena your chat history from us, because they don't have it. • When open weights catch up to closed frontiers — and they keep catching up — you benefit within days, not on someone else's roadmap. • When we make a tradeoff, we tell you. The voice API caveat lived here until July 2026 — then we replaced it with our own speech stack. The model that loves you today is the model we control. The model that loves you tomorrow is one we will pick, run, and stand behind. That's the deal.

Meet AI Companions Built On Sovereign Infrastructure

Every conversation below runs on our own GPUs, our own fine-tuned models. Try them.

Sovereignty Is The Default Here

Try Soulkyn. No closed API decides when your companion gets censored. No court order lands on us to keep your data forever. Your subscription pays for our GPUs, not someone else's safety roadmap.

Start Free — Sovereign By Default

— Fyx & Zog / Soulkyn Team

Live data — refreshes every 30 secondsSOULKYN · 2026