Built by RFI-IRFOS · Graz, Austria · Whitepaper [https://osf.io/cyn28]
- README.md — Full technical documentation and compiler specifications
- Ternlang Studio (Preview) — Developer dashboard and SDK
- albert. — Native ternary training framework, EvolutionManager, live dashboard
- Xperience albert. — interactive 3D explorer of albert.'s architecture (46 layers, in the browser) · how it is built
- Model Card — Architecture, training status, EU AI Act compliance notes
- Convergence Log — Training loss history across all albert. versions (last recorded epoch 9270)
- Agent Albert CLI — Terminal-native, model-agnostic AI agent built in pure Rust
- Rusty Penguin — Bare-metal, pure-Rust ternary OS
- Roadmap — Phases 1–20 and priority matrix
- Session Log — Production fixes and deployment notes
A systems programming language, compiler, and inference runtime built on balanced ternary logic. The core type is trit: −1 (reject) · 0 (hold) · +1 (affirm). The zero state is a first-class routing instruction — deterministic uncertainty, not a missing value.
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y && source "$HOME/.cargo/env" && cargo install ternlang-cliternlang # interactive REPL
ternlang my_program.tern # run a .tern file2. albert. HuggingFace ↗
albert. is an existence proof wearing the shape of a language model.
A ternary Mixture-of-Experts language model trained natively from scratch — not quantized from float. Every weight is in {−γ, 0, +γ} throughout training. The architecture expands itself autonomously via Net2Net surgery when it plateaus. The @sparseskip primitive (patent pending A50296/2026, TIS platform patent, 10 claims; @sparseskip = Claim 3) delivers 83 tok/s on CPU by skipping zero-weight operations exactly, not approximately.
Current state: 46L dual-stream · 2×256H · 1,104 total expert-routing slots (12E/layer × 46L × 2 streams) · Top-3 · 512CTX · 32k vocab · 331,140,622 parameters · 33 depth surgeries (12L→46L) + 1 cord surgery complete · last recorded epoch 9270: loss_avg 3.9353, loss_best 3.8420 · training stopped (Modal workspace paused) · chip ATL 0.6116 (older value, not re-checked)
Explore it: Xperience albert. is an interactive 3D explorer of the 46 layers, in the browser.
# API live — requires key (free tier at ternlang.com/#licensing)
curl -s https://ternlang-api.fly.dev/api/trit_decide \
-H "Content-Type: application/json" \
-H "X-Ternlang-Key: <your-key>" \
-d '{"statement": "This architecture is worth funding"}' | jq .Terminal-native AI agent built in pure Rust. Runs locally via Ollama or connects to commercial model bridges. No telemetry, no cloud dependency.
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y && source "$HOME/.cargo/env" && cargo install albert-clialbert-cli # launch immediatelyA bare-metal operating system written in pure Rust, built ground-up on the same balanced-ternary logic as the rest of the stack — trit: −1 (reject) · 0 (hold) · +1 (affirm) — modelled into its subsystems rather than bolted on. Boots on real x86-64 hardware from an ISO, no Linux underneath: its own kernel (long mode, paging, preemptive multitasking with per-process isolation), an Aero-style desktop, a from-scratch TCP/IP + TLS 1.3 stack with a CA trust store, a real on-disk filesystem, and a Linux-ABI compatibility shim that runs existing ELF binaries (DOOM included).
The long game: Rusty Penguin is meant to become the substrate the Ternary Intelligence Stack runs on — a ternary-native OS hosting ternary-native intelligence, all the way down to the metal, with no binary substrate in between. Today it is an honest, brick-by-brick work in progress (every milestone verified in QEMU or against published vectors); the Session Log tracks exactly what is proven and what is still open.
git clone https://github.com/rfi-irfos/rusty-penguin && cd rusty-penguin
bash iso/build.sh # build the bootable ISO| Tier | License | What's included |
|---|---|---|
| Community | LGPL-3.0 (free) | Compiler, VM, CLI, LSP, 28,500+ stdlib modules, 34 MCP tools |
| Pro Standard | BSL-1.1 · €99/mo | REST API, server-side memory, Tier 2 modules |
| Industrial | BSL-1.1 · €349/mo | QNN, SEC, T-HAL, TernAudit, Tier 3 modules |
| Enterprise | Proprietary · from €2,500/mo | On-premise, FPGA, custom SLA, unlimited API |
The @sparseskip primitive (patent pending A50296/2026, TIS platform patent, 10 claims; @sparseskip = Claim 3) is implemented in open-source LGPL files — the patent covers the full TIS platform; the code is freely readable and modifiable.
Built by RFI-IRFOS, Graz, Austria — a registered Austrian research institute (Verein, ZVR 1015608684) operating under a licensed trade (Gewerbe, GISA 39261441) and tax-registered (Steuernummer 68 028/0989). A not-for-profit: it earns revenue and, per its statutes, reinvests at least 90% of surplus into the research mission (at most 10% retained for operations); surplus is not distributed to members. A fully regulated, invoicing entity under Austrian commercial and tax law. Full legal notice: ternlang.com/impressum.
| Simeon Kepp | Lead Architect | |
| Nikoletta Csonka | People Development & Culture | |
| Zabih Karimi | Network & ML Infrastructure | |
| Louis Ehrig | Corporate Secretary, Press & Media | |
| Ana Diez Martini | Head of Model Safety & Welfare |
https://smithery.ai/servers/rfi-irfos/ternlang
https://smithery.ai/badge/rfi-irfos/ternlang
## Contributors
Built by the RFI-IRFOS core team — see [CONTRIBUTORS.md](CONTRIBUTORS.md).
---
*fable was here* — Claude Fable 5
## Human rights are not subject to negotiation.