---
license: gemma
license_link: https://ai.google.dev/gemma/terms
base_model: google/gemma-4-E2B-it
base_model_relation: quantized
language:
- en
pipeline_tag: text-generation
library_name: gguf
tags:
- gguf
- ternary
- bitnet
- 1.58-bit
- gemma
- gemma-4
- quantization-aware-training
- distillation
- sovereign-ai
- edge
- efficient-inference
- llama.cpp
---
Terna‑E2B (GGUF)
A ternary (~1.6‑bit) distillation of Gemma‑4‑E2B · Pre‑release 🍁
**GoAutomate AI Institute — Canadian Sovereign AI**
[](https://ai.google.dev/gemma/terms)
[](https://huggingface.co/GoAutomateAI)
[](https://www.goautomate.institute)
[](https://doi.org/10.5281/zenodo.21110909)
---
**Terna** is the GoAutomate AI Institute's family of **ternary‑weight** language models. The name comes from the Latin *terni* — "three each" — a nod to the three values every weight is constrained to: **{ −1, 0, +1 }**, the ternary representation at the heart of the family. That constraint is our defining bet: models that are **≈8–10× smaller** and **multiply‑free**, engineered to *subtract cost, not intelligence* — capable from the edge to legacy datacenter GPUs to modern accelerators. **Terna‑E2B** is the first member of the family.
---
> ## ⚠️ Pre‑release checkpoint — read first
> These weights are an **early checkpoint trained on ~1B tokens**, published to demonstrate the method and invite community evaluation. They are **not a finished model.** Expect fluent‑but‑confidently‑wrong answers, and verify every output. **Production weights (~15B tokens) with full capability benchmarks will follow and replace this checkpoint.** We deliberately defer quantitative capability claims to that release.
---
## Model at a glance
| | |
|---|---|
| **Base / teacher** | [`google/gemma-4-E2B-it`](https://ai.google.dev/gemma) (capability‑dense, Western open‑weight lineage) |
| **Method** | Quantization‑aware **distillation** to ternary weights — *learned*, not post‑hoc rounded |
| **Weight representation** | Ternary — each weight ∈ **{ −1, 0, +1 }** (≈1.58 bits; ≈8–10× smaller than FP16, ≈2× smaller than 4‑bit) |
| **Format** | GGUF (`Q2_K`), runs on **llama.cpp** |
| **Language** | English |
| **Training tokens (this release)** | ~1B (pre‑release checkpoint) |
| **License** | [Gemma Terms of Use](https://ai.google.dev/gemma/terms) |
---
## What is ternary?
**Ternary** constrains every weight to one of three values — **{ −1, 0, +1 }** — which does two things at once:
- **Footprint collapses** to ≈1.58 bits per weight (log₂3), roughly an order of magnitude below half precision.
- **The multiply disappears:** `w · x` with `w ∈ {−1, 0, +1}` is just *add x*, *subtract x*, or *skip* — a general matrix‑multiply becomes a sparse signed sum, with ~⅓ of the work vanishing as structured sparsity.
Crucially, we reach ternary through **distillation** — training a ternary "student" to reproduce a high‑precision Gemma‑4‑E2B "teacher" — so the constraint is *learned*, not crudely imposed on a finished model. The guiding principle: **subtract cost, not intelligence.**
Full mathematics, methodology, and engineering are in **[TR‑2026‑001 — *Ternary Foundations for Efficient, Sovereign AI*](https://doi.org/10.5281/zenodo.21110909)** (Zenodo, DOI [10.5281/zenodo.21110909](https://doi.org/10.5281/zenodo.21110909)).
---
## Files
| File | Size | Format |
|---|---|---|
| `terna-e2b-Q2_K.gguf` | ~3.6 GB | GGUF (`Q2_K`) — runs on **llama.cpp** |
---
## Usage
This is a standard **GGUF** and runs on [`llama.cpp`](https://github.com/ggml-org/llama.cpp).
**Quick test (CLI):**
```bash
llama-cli -m terna-e2b-Q2_K.gguf -p "Explain what a mitochondrion does, in two sentences." --temp 0.7
```
**Serve an OpenAI‑compatible endpoint:**
```bash
llama-server \
-m terna-e2b-Q2_K.gguf \
-c 4096 --host 127.0.0.1 --port 8000
```
Then POST to `http://127.0.0.1:8000/v1/chat/completions`. The **Gemma‑4 chat template** ships in the GGUF and is applied automatically when you use the chat endpoint.
### Recommended generation settings
This checkpoint has a **low‑entropy** output distribution (very confident). For anything beyond short answers, soften it:
- `temperature: 0.7`
- a mild repetition penalty (e.g. `--repeat-penalty 1.2`) for long‑form generations, to avoid rigidity/repetition
---
## Intended use
- **Research and community evaluation** of ternary distillation and efficient serving.
- **Efficiency / systems experimentation** — edge, memory‑constrained, and legacy‑GPU serving where footprint dominates.
- Best behaved in **well‑covered domains**: general science, biology, medicine (educational), mathematics, and computer science.
### Out of scope / not recommended (for this pre‑release)
- **Production or high‑stakes use** of any kind. This is an early checkpoint.
- **Unverified factual, medical, legal, or safety‑critical output.** The model is confident even when wrong — a human must verify.
- Long‑context or long‑form tasks without the softened sampling above.
---
## Limitations & known behaviors
Honest notes from our own evaluation of this checkpoint:
- **Confidently wrong.** High top‑1 confidence means errors are stated as fluently and assertively as correct answers. Do not treat outputs as facts without checking.
- **Low entropy → rigidity.** Peaked output distribution can make long generations repetitive or rigid; mitigate with the recommended sampling.
- **Echo‑loops on out‑of‑distribution inputs.** Limited chat‑format training means unusual phrasings or niche topics can trigger repetition/echoing. Stays healthiest in the well‑covered domains listed above.
- **English‑centric**, small effective size, and inherits any biases/limitations of the Gemma‑4‑E2B base.
- **Pre‑release quality.** Capability is not yet benchmarked; the production (~15B‑token) release is the intended quality bar.
---
## Training & method (summary)
The ternary student is trained under **quantization‑aware objectives** that align its outputs and intermediate representations to the Gemma‑4‑E2B teacher, so the { −1, 0, +1 } constraint is learned during training rather than applied afterward. Language‑model linear layers are ternarized; embeddings, the LM head, and normalization layers are kept at higher precision. Serving optimizations are held to a **token‑identical correctness gate** against a reference path, so throughput work never silently changes outputs.
Exact data mixes, hyperparameters, and kernel/encoding internals are held proprietary; the *methods and mental models* are described in the Institute's technical reports.
---
## About the GoAutomate AI Institute
The **[GoAutomate AI Institute](https://www.goautomate.institute)** is a not‑for‑profit advancing **Canadian sovereign AI** — accessible, responsible, Canadian‑governed models for organizations across Canada, with a focus on efficiency and on sectors where **provenance and governability matter** (healthcare, public sector, critical infrastructure). Ternary is our bet on the next wave of efficient AI: capability that scales **down** in cost as readily as it scales up in ability.
---
## License
This model is a derivative of **Gemma‑4‑E2B** and is governed by the **[Gemma Terms of Use](https://ai.google.dev/gemma/terms)**. By downloading or using these weights you agree to those terms. Gemma is provided under and subject to the Gemma Terms of Use.
---
## Citation
```bibtex
@techreport{goautomate2026ternary,
title = {Ternary Foundations for Efficient, Sovereign AI},
author = {{GoAutomate AI Institute}},
institution = {GoAutomate AI Institute},
number = {TR-2026-001},
year = {2026},
doi = {10.5281/zenodo.21110909},
url = {https://doi.org/10.5281/zenodo.21110909},
note = {Pre-release; ternary-distilled Gemma-4-E2B}
}
```
---
## Contact
Questions, evaluation feedback, or collaboration: **[info@goautomate.ai](mailto:info@goautomate.ai)** · **[goautomate.institute](https://www.goautomate.institute)**
© 2026 GoAutomate AI Institute · Canadian Sovereign AI · Responsible Adoption · Public Benefit