CYBER-FROST-3.8-HGN

An independently converted native HGN v2 derivative of Blackfrost-AI/CYBER-FROST-3.8-BF16, published by Quat3rnion. This is not HT43, not a BF16 checkpoint, and not an official Blackfrost-AI or Qwen release.

Source and attribution

The exact conversion source is pinned to revision 3e30d25ddb0e802389d62e55e9ff31ad05f0c96e. Blackfrost-AI supplies the security-adapted model. Its pinned model card identifies Qwen/Qwen3.8-Flash-Next at revision de4b8e4d43b917e7706784d8bb445c9af86a3540 as the foundation and describes security fine-tuning, a merged BF16 checkpoint, and a proprietary behavioral stage. Tokenizer and processor lineage comes from that Qwen foundation; the chat template is Blackfrost-specific.

The source's Qwen Community License 1.0 is reproduced verbatim in LICENSE, including its copyright notice and commercial-use conditions. Read it before use or redistribution. The source card explicitly says mixed-source training-data provenance and licensing review remains in progress; conversion does not resolve that limitation. Source claims, benchmark results, and BF16 deployment performance do not automatically transfer to this quantized derivative.

Conversion and files

The actual encoder was the original schema-1 CPU bf16-to-hgn tool, version 0.1.0, with seed 0, two CPU threads, and HT beam 128. This checkpoint was not GPU-encoded. The conversion manifest records HGN specification version 1.1.0 and specification commit 79e75fd4ae7642478823525b0a3212951f584af5. The current converter/verifier repository is provided for inspection and verification; it is not the historical encoder revision used for this artifact.

Configuration, tokenizer, chat-template, and image/video processor assets are included. Keep the three HGN files together with the packaged assets. Do not substitute stock or unrelated ngram/vision sidecars. Original BF16 shards, credentials, machine paths, and internal run logs are not distributed.

Verification evidence

The saved full CPU and HIP verification runs both completed successfully. They checked 1,658 source tensors, 2,349 output tensors, and 118,785,459,868 payload bytes. Their 1,531 per-tensor quantization-error records are bit-identical between CPU and GPU. These are error-metric records, not 1,531 verification failures. The HIP verifier evaluated the CPU-encoded artifact; it did not encode it.

The manifest is schema 1 and does not contain embedded provenance attestation. The pinned source and encoder description above are recorded conversion provenance, not a claim of schema-2 attestation. CPU/GPU metric agreement establishes reproducibility of those checks, not guaranteed quantization quality, BF16 equivalence, comprehensive capability, or production readiness.

Exercised Halogen profile and template requirement

Saved inference evidence passed on Halogen API/engine 0.17.5, with a 32,768-token context and one slot, using this checkpoint and its own vision/ngram sidecars, without loading NPU models:

• Arithmetic: 37 * 43 returned 1591.
• Exact-word instruction: returned blackfrost.
• Structured tool use: proposed get_weather with {"city": "Boston"}.
• Vision: identified a solid red test image as Red.

These are narrow smoke tests, not benchmark or multimodal-quality qualification. Video understanding, longer contexts, concurrent serving, sustained agent behavior, and broad cybersecurity performance were not established by these smokes.

The packaged Blackfrost chat template forces thinking. Halogen's template probe found identical rendering for enable_thinking=false and true: enable_thinking, reasoning_effort="none", and HALOGEN_ENABLE_THINKING=0 do not disable its thinking. The exercised run used the explicitly user-approved HALOGEN_TEMPLATE_UNCHECKED=1 override. Halogen 0.17.5 requires that intentional opt-in to serve this unchanged template; it bypasses a failed thinking-control probe, not a quality or safety check. Do not assume non-thinking requests are respected. Keep the original template for the documented behavior, and explicitly authorize any override in your own deployment.

Intended use and limitations

Use only for professional security work within explicit authorization. The model cannot establish ownership, consent, legal scope, or tool permissions. Enforce authorization, isolation, least privilege, logging, human review, and credential protection externally. Reduced refusal behavior is not an access-control mechanism or safety guarantee.

Generated analysis, code, commands, and tool calls can be wrong, incomplete, unsafe, or fabricated. Independently review outputs and execute only within approved environments. Quantization can change behavior and quality; no comprehensive quality comparison with the source BF16 checkpoint is claimed. This derivative is provided without warranties, subject to the accompanying upstream license.