Public work

What we publish in the open.

Technical output anyone can open without asking us: open-weight model builds and the datasets collection on Hugging Face, ten incident runbooks, the rules of engagement we sign, and a research paper. Each item says what it does and does not show.

Hugging Face snapshot 6 September 20264 public model repositories0 public datasets

Datasets

The datasets collection.

No public datasets were visible in the Qompute AI Hugging Face organization on the retrieval date. This page will list them, with their license and intended use, once one is published and approved for reference.

Qompute AI datasets on Hugging Face

The collection is separate from qore, from proprietary models, and from any claim of original model training. Datasets appear here only with a stated license, provenance, and intended use.

Model builds

GGUF conversions of open-weight models.

Conversions of third-party models under their original licenses, usable by any llama.cpp-based runtime. They are not models trained by Qompute AI and they are not a qore download.

Public model

Gemma 4 E2B (instruction-tuned), GGUF

A small multimodal open-weight model converted to GGUF, with vision projector files, for local inference on modest hardware.

Shows: That Qompute AI can convert and publish a current open-weight model in the format qore's inference runtime consumes, including the projector files needed for image input.
Does not show: It does not show model quality, Qompute AI training capability, or any qore feature. Accuracy is the base model's.
Useful for: Developers evaluating local multimodal inference; qore preview testers choosing a starter model.
License: gemma (the base model's license; see the repository)
Files: bf16, q8_0, q4_k_m, mmproj bf16 and q8_0 · 5 GGUF files
Access: Public, not gated

Snapshot retrieved 6 September 2026QomputeAI/QomputeAI-Google-Gemma-4-E2B-it-GGUF

Open on Hugging Face

Public model

Gemma 4 E4B (instruction-tuned), GGUF

A mid-size open-weight model converted to GGUF in three precisions.

Shows: Repeatable conversion across the Gemma 4 family with standard llama.cpp quantization types.
Does not show: It does not show model quality or any qore feature. The repository name lacks the GGUF suffix used by the others; the files are GGUF.
Useful for: Developers and preview testers with 16 GB or more of unified memory.
License: gemma (the base model's license; see the repository)
Files: bf16, q8_0, q4_k_m · 3 GGUF files
Access: Public, not gated

Snapshot retrieved 6 September 2026QomputeAI/QomputeAI-Google-Gemma-4-E4B-it

Open on Hugging Face

Public model

Gemma 4 12B (instruction-tuned), GGUF

A 12-billion-parameter open-weight model converted to GGUF in three precisions, with a longer model card.

Shows: A larger conversion with a written model card covering hardware guidance and a llama.cpp run example.
Does not show: The model card describes a custom quantization scheme; the published files use standard llama.cpp quantization types (bf16, q8_0, q4_k_m). This site describes only what the files show. It does not show any qore feature.
Useful for: Developers with 12 GB or more of GPU or unified memory.
License: gemma (the base model's license; see the repository)
Files: bf16, q8_0, q4_k_m · 3 GGUF files
Access: Public, not gated

Snapshot retrieved 6 September 2026QomputeAI/QomputeAI-Google-Gemma-4-12B-it-GGUF

Open on Hugging Face

Public model

DiffusionGemma 26B A4B (instruction-tuned), GGUF

A large open-weight model converted to a single bf16 GGUF file.

Shows: Conversion of a large model; the file is about 50 GB and is intended for high-memory hosts.
Does not show: One precision only; no quantized variant; minimal model card. It does not show any qore feature.
Useful for: Researchers and operators with very high-memory hardware.
License: gemma (the base model's license; see the repository)
Files: bf16 · 1 GGUF file
Access: Public, not gated

Snapshot retrieved 6 September 2026QomputeAI/QomputeAI-Google-DiffusionGemma-26B-A4B-it-GGUF

Open on Hugging Face

Model names include the original publisher's name because that is how the base models are identified. No endorsement by or relationship with the model publisher or with Hugging Face is implied; Hugging Face is the hosting platform. Facts above come from a snapshot of public metadata retrieved 6 September 2026; download counts are not shown because they change.

Method, published

How we work, in writing.

10 incident runbooks

First fifteen minutes, first hour, evidence, containment, durable controls, and closure for the incidents AI systems have. Free to use inside your organization. Read them.

Rules of engagement

The template both sides sign before any adversarial module runs. Read the template.

Research

One published paper, at its actual scale.

Hinch, 2024. Revolutionizing AI with Quaternion Algebra: A Leap in Neural Network Efficiency. 98.04% test accuracy on MNIST with 49,170 trainable parameters, test loss 0.083, training under two seconds. The cited dense baseline of roughly 200,000 parameters is the author's comparison, not an independent one.

MNIST is a toy benchmark. This result says nothing about production accuracy, language models, or any customer workload.
Extension to convolutional, attention, or embedding architectures is ongoing work, not a result.

Research is kept separate from product capability and from client work; nothing here is product evidence or a delivery commitment. Other internal experiments are not published until they have a measured result on a defined corpus.