Models
Published checkpoints and lineage
Each model states its architecture, training data, evaluation scope and exact publication status. Native and CoderSec models are independent and must not be treated as one architecture family.
| Model | Lineage | Parameters | Tokenizer | Status | Artifact |
|---|---|---|---|---|---|
| XRFM Native 25M | Original, random initialization | 25,200,000 | T4 32K BPE | Published | Hugging Face |
| XRFM Native 85M | Original, random initialization | 86,446,464 | T4 32K BPE | Published | Hugging Face |
| XRFM CoderSec | Qwen3-1.7B base + QLoRA | ~1.7B base | Qwen tokenizer | Not published | See status page |
Native weights
Released under Apache-2.0 metadata. Repository source remains MIT-licensed; the weight license is artifact-specific.
CoderSec attribution
Base model Qwen/Qwen3-1.7B revision 70d244cc, Apache-2.0. No adapter or merged weights are public.
No hidden lineage
Native models loaded no third-party pretrained weights. CoderSec did, and says so on every public surface.