GoAutomate AI Institute
Canadian Sovereign AI, built for public benefit đ
**Accessible ¡ Responsible ¡ Canadianâgoverned**
[](https://www.goautomate.institute)
[](https://doi.org/10.5281/zenodo.21110909)
[](mailto:info@goautomate.ai)
Advancing efficient, sovereign artificial intelligence for Canada â and sharing the methods with the world.
---
## đď¸ Who we are
The **GoAutomate AI Institute** is a notâforâprofit advancing **Canadian sovereign artificial intelligence** â AI models, tools, and research designed for Canadian organizations and aligned with Canadian values, ethics, and governance.
We don't believe Canada has to win only the race toward everâlarger frontier models. **Purposeâbuilt, efficient, Canadianâgoverned models** can deliver real outcomes in real environments â hospitals, health authorities, public institutions, businesses, researchers, and notâforâprofits â while keeping people in control and accountable.
This organization is where we publish our **openâweight models, inference research, and technical reports.**
### At a glance
| | |
|---|---|
| **Mission** | Canadian sovereign AI, for public benefit |
| **Model family** | **Terna** â ternaryâweight models (*Latin* terni, "three each") |
| **First release** | **TernaâE2B** â ternary distillation of Gemmaâ4âE2B *(preârelease)* |
| **Format** | GGUF ¡ runs on `llama.cpp` |
| **Flagship report** | TRâ2026â001 ¡ [DOI 10.5281/zenodo.21110909](https://doi.org/10.5281/zenodo.21110909) |
| **Structure** | Notâforâprofit ¡ open weights |
---
## đŹ Our research bet: ternary
Modern AI isn't limited by intelligence â it's limited by **the cost of storing and moving it**. Most of the cost of running a language model is the memory bandwidth spent streaming billions of weights for every token, and the arithmetic of multiplying them.
**Ternary** representation constrains every weight to one of three values â **{ â1, 0, +1 }** â and attacks both costs at once:
| â 8â10Ă smaller | Multiplyâfree | Capability preserved |
|:---|:---|:---|
| Each weight carries â **1.58 bits** (logâ3) instead of 16 â roughly an order of magnitude below half precision, and about half the size of 4âbit quantization. | A weight in { â1, 0, +1 } turns the expensive multiply into **add**, **subtract**, or **skip**. Roughly a third of the work disappears as structured sparsity. | Reached through **distillation** â a ternary student trained to reproduce a highâprecision teacher, so the constraint is *learned*, not imposed after the fact. *Subtract cost, not intelligence.* |
The payoff spans hardware generations: ternary's footprint lets capable models **fit on memoryâconstrained edge devices** and **revive legacy datacenter GPUs** that modern models had outgrown â extending the useful life of existing silicon â while on modern accelerators the same savings mean **far more concurrent users per device**. It's an environmental story as much as a performance one: less energy per token, fewer accelerators per unit of served intelligence, and a slower path to eâwaste.
---
## â Featured release â TernaâE2B
[](https://ai.google.dev/gemma/terms)
**TernaâE2B** is our first release, and the first in the **Terna** family of ternaryâweight models (*Terna*, from Latin *terni*, "three each"). It is a ternary (â1.6âbit) distillation of **Gemmaâ4âE2B**, a capabilityâdense model with a **transparent, Western openâweight lineage**, distributed as **GGUF** for efficient local and datacenter serving.
> â ď¸ **Preârelease checkpoint.** The published weights are an early checkpoint (~1B training tokens), shared to demonstrate the method and invite community evaluation. **Production weights (~15B tokens) with full capability benchmarks will follow and replace them.** We defer quantitative capability claims to that release rather than overâstate a preârelease checkpoint.
*Browse the model repositories in this organization for files, quantizations, and usage.*
---
## đŻ Why Gemmaâ4 as the base
- **Exceptional intelligenceâperâFLOP.** An already capabilityâdense family; ternary distillation compounds that efficiency.
- **Elastic architecture.** The E2B ("effectiveâ2B") design fits edge and multiâtier deployment â and ternary lowers the floor of that range further.
- **Provenance you can defend.** For regulated and sovereigntyâsensitive sectors â healthcare, public sector, critical infrastructure â the *provenance* of a model's knowledge is a governance requirement, not a preference. A transparent, auditable, Western openâweight base gives the resulting ternary model a clear, documentable lineage. **Efficiency should not come at the cost of provenance.**
---
## âď¸ An engine built to spend ternary's footprint on throughput
Ternary's memory savings are only as valuable as a serving stack's ability to convert them into useful work. We maintain an inference engine (a fork of the openâsource `llama.cpp` lineage) tuned so footprint becomes **concurrency and throughput**:
- **Footprint â concurrent contexts** â small weights leave far more device memory for KV cache, so one accelerator serves many more simultaneous sessions.
- **Hardwareâbest matmul dispatch** â the engine picks the path that fits the device by batch size, so capabilityâdense models serve efficiently even on GPUs that lack dedicated matrixâmultiply units.
- **Deviceâside sampling** and **sharedâprefix / crossâsession cache reuse** â agentic workloads pay for a large shared system prompt once, not once per user.
We validate this across **deliberately different accelerator families** â legacy datacenter GPUs and earlierâgeneration tensorâcore GPUs â to prove the ternaryâplusâengine approach spans hardware generations. Every serving optimization is held to a **tokenâidentical correctness gate** before it ships: throughput work never silently degrades output.
---
## đ Research
[](https://doi.org/10.5281/zenodo.21110909)
- **[TRâ2026â001 â *Ternary Foundations for Efficient, Sovereign AI*](https://doi.org/10.5281/zenodo.21110909)** â the mathematics of ternary, why Gemmaâ4, the engine that turns footprint into throughput, and the implications for legacy and modern hardware. *(Zenodo ¡ CCâBYâ4.0)*
- **TRâ2026â002 â *QuantizationâAware Training and Ternary Weights: Method vs. Representation*** â clears the "QAT vs ternary" category error and places both on one efficiency curve.
Further reports â on serving engines across accelerator families, multiâGPU parallelism for heterogeneous fleets, sharedâcontext concurrency, ternary at larger scales, and governable agentic orchestration â are on the Institute's roadmap.
---
## đ¤ What we stand for
| | |
|---|---|
| đ **Sovereignty** | Reduce dependency on foreignâcontrolled AI models and infrastructure; support Canadian laws, governance, and interests. |
| đĄď¸ **Responsibility** | Steerable, explainable, accountable, secure systems that keep humans at the helm. |
| đ **Accessibility** | Capable models, tools, and documentation available to organizations across Canada. |
| ⥠**Efficiency** | Capability that scales **down** in cost as readily as it scales up in ability. |
---
## đŹ Get involved
The GoAutomate AI Institute welcomes collaboration with **Canadian organizations, healthcare systems, public institutions, researchers, policymakers, and technology leaders** who share our commitment to responsible and accessible AI.
[](https://www.goautomate.institute)
[](mailto:info@goautomate.ai)
---
Š 2026 GoAutomate AI Institute ¡ Canadian Sovereign AI ¡ Responsible Adoption ¡ Public Benefit