PYN-S
32B dense
The efficient tier. A capable assistant that runs on a single GPU node, suited to focused Q&A and internal knowledge work.
Sovereign AI
Pyon Labs trains private language models on your institution's own corpus and deploys them where you decide.
Sovereign AI is built for banks and financial institutions, professional bodies and educators, and regulated enterprises.
Trusted institutions should own their AI, not rent it.
- Your model, your data
- The model is trained on your own corpus, such as policies, curricula, procedures and product documentation, and nothing leaves your control.
- Ownership stays in-house
- Your corpus, guardrails and weights stay yours. Escrowed licensing means you own the model rather than rent it.
- Aligned to your standards
- Every build ships with an evaluation harness scored against your own material, plus a red-team review.
- Pyon supplies the engine
- Pyon provides the corpus engineering, training, alignment and deployment infrastructure as a product rather than a bespoke project.
The Sovereign Series
Four model tiers share one promise: frontier-class architecture, trained for your institution. Larger tiers use sparse Mixture-of-Experts, with the scale of a 700B+ parameter model, the serving cost of its active parameters and pre-training on over 10 trillion tokens.
PYN-M
110B MoE · 12B active
The workhorse. Full co-pilot capability with strong reasoning, at a serving cost that makes organisation-wide rollout practical.
PYN-L
355B MoE · 32B active
The flagship assistant tier, built for the hardest analytical workflows and the deepest domain corpora.
PYN-Max
700B+ MoE · 40B active
Frontier-class sovereign capability on a multi-node cluster, for institutions that want the ceiling.
Every build includes corpus engineering, fine-tuning and alignment, an evaluation harness scored against your own standards, and red-team review. Multilingual training is available, including Arabic and English.
Two ways to run it.
Sovereign on-prem
The model runs inside your own data centres, air-gapped if required. Its weights are licensed to you and held in escrow, your data never leaves your systems, and internal use is unlimited.
Pyon-managed
Pyon Labs runs a dedicated, single-tenant instance for you, with regional residency options, encryption throughout and no training on your users' data. You own the model and we carry the pager.
From corpus to co-pilot
Scoping starts with a two-week corpus audit and a working prototype on your own material.
Corpus audit
A two-week scoping engagement produces a data-readiness report, a fixed quote and a working prototype on a sample of your material.
Training and alignment
Your Sovereign Series model is trained on your prepared corpus and aligned to your voice and rules.
Evaluation and red-team
The model is scored against your own standards and stress-tested before anyone inside your organisation sees it.
Deployment
The model goes live on your infrastructure, air-gapped if required, or as a dedicated managed instance run by Pyon Labs.