Qovaryx 350M scratch base
Random-initialized weights. Training is required before this artifact can perform a useful language or classification task. The size in the repository name is a family label; it is not a benchmark result.
Status — October 3, 2026: Qovaryx is on the back burner. Desktop downloads are paused; there is no current public installer, purchase, or trial. These repositories remain as research references. Published weights are separate from desktop product availability.
Configuration
| Field | Published value |
|---|---|
| Architecture | QovaryxForCausalLM |
| Width | 1024 |
| Decoder layers | 24 |
| Attention heads | 16 |
| Feed-forward width | 2816 |
| Vocabulary size | 20,242 |
| Configured maximum sequence length | 2,048 |
These values come from config.json. They describe the supplied configuration, not measured training throughput, hardware compatibility, or model capability. Optional architecture mechanisms require compatible code, configuration, and training; their presence does not establish learned performance.
Repository format
This release contains pytorch_model.bin, a tokenizer, and custom Transformers wrapper files: _qovaryx_decoder.py, configuration_qovaryx.py, and modeling_qovaryx.py. train_quickstart.py is a training reference.
The recorded loading interface is:
from transformers import AutoConfig, AutoModelForCausalLM
repo = "tjarvis91/qovaryx-350m-scratch-base"
# Review the repository's custom Python code before enabling trust_remote_code.
config = AutoConfig.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)
This documents the bundled wrapper interface; it was not rerun during this documentation update. Choose the device, precision, dependencies, and training setup for your own environment. No specific GPU-memory fit or universal CPU support is certified here.
Evaluation and lineage
No task-performance result is claimed for these random weights. Results from separately trained specialist or Cluster Shell experiments cannot be attributed to this untrained checkpoint. Historical claims about frontier-level cognition, general hardware fit, or the quality of optional backends have been removed.
The trained Q-* specialist releases are separate checkpoints with their own configurations, tokenizers, dependencies, and evaluation limits. A scratch-base repository is not evidence that a particular specialist was trained directly from these exact uploaded weights.
License
Apache-2.0, as recorded in the existing model metadata. Review the supplied code and any third-party training data or dependencies separately.
Project and support
J.E. Herizon LLC, doing business as JE Horizon.