CoolFace
Modelpublic

firehazardlabs/vykon-htma-2

sourceHugging Faceapache-2.0updated 10d agoView on Hugging Face
0likes21downloads
Model Card

Vykon HTMA-2 — development candidate

Stage 2 step 1466 is the best available development candidate. Both evaluated Stage 2 checkpoints failed the absolute release gates. This model is not clinically validated or deployment-ready. Known unsafe recommendations, accuracy failures, invalid schema and truncation remain unresolved.

Public artifact publication

The repository owner authorized public model downloads on 14 September 2026. This includes all model branches. Historical snapshot cards describe the original private export; their visibility statement is superseded by this notice. Artifact publication does not establish clinical safety or authorize clinical use.

Download an exact variant

The default main branch contains this index only. Select an immutable model revision below when downloading or serving; do not load main as model weights.

VariantImmutable revisionArtifact size
BF16`df0fdba42585df21b052fec262a1944a42134ac1`55.59 GB
FP8`bfefd3ab6d256b1c12a60b88012babf88f7c6a5c`31.24 GB
Mixed NVFP4/FP8`3c42e688856fbc091bbbf34988f878aef2e887a5`21.94 GB

Both quantized variants were produced independently from the same merged BF16 parent. They passed tensor/configuration checks and an eight-prompt numerical smoke test in vLLM 0.28.0 on an RTX PRO 6000 Blackwell Server GPU. These are artifact engineering checks, not clinical evaluations. H100 execution, long-context behavior, concurrency and production acceptance remain unqualified.

Per-file SHA-256 hashes, source-base revision and parent lineage are included in each variant's artifact-manifest.json. Calibration dataset files, raw case files, credentials and private logs are not included. This does not establish that the model is free of memorized training content.