AURIX is a compact, open-weight AI platform engineered for absolute privacy — offline inference via llama.cpp, user-controlled memory, and a modular tool engine. No data leaves your device. Ever.
Every architectural decision in AURIX traces back to six pillars. These aren't features — they're constraints that shape everything.
All inference runs via llama.cpp GGUF on your machine. No API keys. No remote endpoints. No hidden network calls. Air-gap verified.
~3B dense parameters with GQA and SwiGLU for maximum quality at minimum footprint. Runs on 8GB laptops. Fits in 2.5GB VRAM.
Your preferences are stored in inspectable SQLite tables — never baked into model weights. Delete anything, anytime, with cryptographic certainty.
Every tool call passes through a 5-stage pipeline with Pydantic schema validation and human-in-the-loop approval gates.
Vision, speech, and document modules plug in independently. The base engine works perfectly without any of them installed.
Apache 2.0 licensed. Every benchmark is reproducible. Every quantization level is documented. Full transparency, full auditability.
AURIX separates concerns across four independent layers. Each can be swapped, upgraded, or removed without breaking the others. The base model uses a ~3B parameter dense transformer with Grouped-Query Attention for efficient long-context processing.
RMSNorm for training stability · SwiGLU activation for quality · RoPE positional encoding for extensible context · GQA (8 KV heads / 32 query heads) for memory efficiency during generation.
Powered by llama.cpp with native GGUF format. Supports Metal (Apple Silicon), CUDA (NVIDIA), Vulkan (AMD/Intel), and pure CPU AVX-512/AVX2 paths — all from a single model binary.
Each profile tunes the routing layer, verification strictness, and tool availability. Switch instantly based on your task context.
Adjust the parameters to see real-time memory, speed, and compatibility estimates for your hardware configuration.
AURIX exposes http://127.0.0.1:8080/v1 — connect any IDE plugin, local script, or desktop app with zero cloud API keys.
from openai import OpenAI
# Connect to your local AURIX instance
client = OpenAI(
base_url="http://127.0.0.1:8080/v1",
api_key="not-needed"
)
response = client.chat.completions.create(
model="AURIX-3B-Instruct",
messages=[
{"role": "system", "content": "You are AURIX."},
{"role": "user", "content": "Summarize this repo."}
],
temperature=0.2,
stream=True
)Your data is yours. AURIX uses a strictly decoupled memory system where neural weights are never modified by user interactions.
Active conversation window. Lives in KV-cache during session, completely destroyed when window closes or context rotates. No persistence. No traces.
Coding rules, tone settings, tool defaults. Stored in ~/.aurix/preferences.json. Fully inspectable, exportable, and deletable. Never touches model weights.
Indexed local documents stored in SQLite vector tables. Embedded via local bge-micro model. Supports cosine similarity retrieval with passage-level citations.
One toggle disables all vector indexing, preference saving, and conversation logging. The session becomes fully ephemeral — as if it never happened.
Every tool call follows a strict 5-stage pipeline. Destructive operations require explicit human approval before execution.
pytest tests/test_auth.py in your sandboxed workspace. This requires authorization:
sandbox_runpytest tests/test_auth.py -vEach modality is an independent, downloadable module. The base 3B engine works perfectly without any of them.
SigLIP/Moondream GGUF plugin. Understand screenshots, diagrams, charts. Runs independently of text engine.
Local Whisper model integration. Transcribe audio directly to text for hands-free interaction. Zero cloud STT.
Piper engine for natural text-to-speech output. Multiple voice models. Fully offline, responsive.
Surya-style layout analysis for scanned PDFs. Extract structured text from images. Integrates with RAG pipeline.
Same design language, different capability tiers. All GGUF-compatible, all locally executable.
AURIX doesn't trust the model for everything. When the task is computational, a deterministic tool handles it. When the task requires citations, retrieval provides grounding.
| Capability | Approach | Status |
|---|---|---|
| Arithmetic | Tool dispatch (not generative) | ✓ Verified |
| Code output | Syntax-validated structured JSON | ✓ Enforced |
| Document claims | Passage-level citation required | ✓ Grounded |
| Uncertainty | Calibrated "I don't know" responses | ✓ Trained |
Every metric is tied to specific hardware, quantization, and context settings. Held-out evaluation sets prevent contamination. Full transparency on methodology.
| Benchmark | Q4_K_M Score | Status |
|---|---|---|
| MMLU (5-shot) | Target ≥ 58% | ◎ In Progress |
| HumanEval | Target ≥ 52% | ◎ In Progress |
| TruthfulQA | Target ≥ 62% | ◎ In Progress |
| Perplexity Delta vs FP16 | ≤ 4% | ✓ Baseline |
From architecture validation through community release. Each phase has concrete deliverables and acceptance criteria.
Download AURIX, run it offline, keep your data. It's that simple. Open-weight, reproducible, yours.