Official Qwen3.8-27B — hf2q Q4_K_M + optional PhantomKV

Q4_K_M conversion of the official Qwen/Qwen3.8-27B checkpoint at 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. The model is the unmodified official checkpoint: no fine-tuning, ablation, or baked-in runtime steering. Quantization changes its numeric representation; it is not byte-identical to the upstream safetensors.

Files and provenance

• Qwen3.8-27B-Q4_K_M.gguf: 16,810,714,976 bytes, SHA-256 cf0eb0103c1ffe7b3702670aa794a669bafbe328b95e892268305dd87c2c30d7.
• Qwen3.8-27B-Q4_K_M.gguf.receipt.json: genuine adjacent conversion receipt from hf2q 0.1.26 commit 9c42dc1fc19b4dbe94c63c6c4bfef7f990c8f856. All source, converter, and output-hash fields are preserved; only output.path is made repository-relative. Native receipt binding deliberately ignores that portable path, not the identity fields.
• Qwen3.8-27B-phantom-kv.gguf: optional, separate runtime-only trained K/V bank, 17,303,808 bytes, SHA-256 e33df498ad4abe47b7d98bf0228dc96cd629db15ec37574c2514ef88f4656ffb. It does not replace or modify the base model weights. It was trained from the same official checkpoint for 400 steps with seed 1337 using the phantom-kv method/toolchain. The direct-K/V F32 container holds 132 slots across 16 full-attention layers: 3, 7, 11, 15, 19, 23, 27, 31, 35, 39, 43, 47, 51, 55, 59, 63. Unbinding the bank and starting fresh runtime state removes the graft.
• manifest.json records checkpoint identity, artifacts, runtime dependencies, and licensing. SHA256SUMS covers all release files except itself. Raw training prompts, datasets, logs, and checkpoints are not distributed.

The graft concept, container, and producer/training method are credited to lordx64 and phantom-kv contributors (MIT); hf2q supplies the conversion and independent native reader/runtime integration. See NOTICE and PHANTOM-KV-LICENSE.

Optional GLP dependency — not redistributed

GLP stays upstream and is not included in this repository. The external GLP for the exact same official source revision is pinned to:

Qwen3.8-27B-abliterated-cyber-GLP-49-L10-58-a1.gguf

SHA-256: 93c0742d597a77f9aea67298ef0abe06675da69bf8f7fa46c79c0365728f31d2; alpha: 1. Its actual file/license is MIT while the upstream repository metadata says Apache-2.0; this discrepancy is noted rather than silently resolved or relicensed. Consult the upstream file and repository terms independently. GLP, GCD, and PhantomKV are runtime controls, not model-weight modifications.

Limitations and licensing

Historical MPS reference results (49/60 to 1/60 refusals; 0/20 harmless refusals) are limited probe results on that reference path, not a native-runtime quality qualification, coding benchmark, or general capability/safety guarantee. Native coding checks are a separate release gate; this card makes no passing claim. Grafts and runtime steering can degrade capability and change behavior. Evaluate the exact combined runtime configuration for your workload.

The official model retains its upstream Apache-2.0 license and Alibaba Cloud attribution, supplied unchanged as LICENSE. The authorized operator-owned derived bank is distributed under Apache-2.0; the producer's MIT license and attribution remain preserved in PHANTOM-KV-LICENSE and NOTICE. No upstream GLP is redistributed or relicensed here. This is an independent conversion/runtime-artifact release, not an endorsement by Qwen, Alibaba Cloud, or the producer authors.