Yaqgu is an independent research lab engineering foundational sequence architectures. By decoupling memory scaling from sequence length, our Ypru foundation model family delivers infinite context and high-throughput edge execution with zero KV-cache memory explosions.
The Yaqgu Ecosystem
Engineered as a clean separation between foundational machine learning research and edge deployment systems.
Autonomous Intelligence & Edge Runtimes
Yaqgu is the parent systems engineering company. We develop privacy-first in-browser agent frameworks, local WebGPU execution layers, and enterprise serverless inference orchestration designed to eliminate cloud GPU hosting premiums.
Sample-Efficient Linear Sequence Recurrence
The core intelligence series powering Yaqgu systems. Ypru is an entirely sovereign base model family combining Taylor-expansion linear attention with dynamic neural memory gating to achieve linear sequence scaling with strictly constant VRAM.
Zero-OOM Hardware Scaling
Audited on NVIDIA H100 SXM5 compute infrastructure under real inference execution.
Pre-Training Volume
100,000,000
Permissive Open Tokens
Convergence Delta
479.76 → 4.23
-99.1% Error Reduction
VRAM at 32k Context
424.64 MB
0 MB Growth vs 1k Context
Optimization Engine
Muon + AdamW
Newton-Schulz Spectral Loop
Open Science & Next Milestones
Yaqgu commits to open research, verifiable model weights, and transparent benchmark publishing.
Pre-trained exclusively on gold-standard educational corpora: FineWeb-Edu (50%), Open-Web-Math (25%), and CodeParrot-Clean (25%). Free of proprietary distillation contamination or ToS conflicts.
Currently scaling the Ypru architecture to a 1.0B parameter model trained across 20 Billion tokens using Cloud TPU Pods to establish an open benchmark for constant-memory linear attention models.
Model weights are released for free academic and personal evaluation under the Ypru Community License, with custom commercial licenses available for enterprise production integration.
Load Ypru-100M via Hugging Face
# Load model directly using standard Transformers library
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("yaqgu/ypru-100m", trust_remote_code=True)
model.eval()
# Executes with O(1) fixed recurrent memory
output = model(input_tensor)