Downloads · 30 days
24
7% of all-time downloads
1105s110/Qwen3.8-27B-Blackfrost-Abliterated-SlotTuned-15G-GGUF
Qwen3.8-27B-Blackfrost-Abliterated-SlotTuned-15G-GGUF is a machine learning model from 1105s110. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for gguf. The card lists the license as apache-2.0.
TL;DR: The 15 GiB tier of our Blackfrost-lineage lineup. A hand-tuned per-slot mixed quant: sensitive tensor classes pinned high, bulk at q4K. KLD 0.0387 vs the BF16 golden — better fidelity than our previous-gen 16.9…
Downloads · 30 days
24
7% of all-time downloads
All-time downloads
335
Public
Repo size
16.1 GB
Likes
0
Public
Click a slice to open those files.
.gguf16.1 GB · 100%
From the Hugging Face model README
TL;DR: The 15 GiB tier of our Blackfrost-lineage lineup. A hand-tuned per-slot mixed quant: sensitive tensor classes pinned high, bulk at q4_K. KLD 0.0387 vs the BF16 golden — better fidelity than our previous-gen 16.9G quant — at HardBattery 14/17, the same score as both our 9.8 GiB IQ3_XXS and the bigger daily driver. All numbers measured locally on the same harness (RTX 5090, llama.cpp b10930).
中文摘要:我们 Blackfrost 血统产品线的 15G 档。手工槽位调优混合量化(敏感槽位钉高、主体 q4_K):对 BF16 黄金基线 KLD 0.0387——优于上代 16.9G 档——HardBattery 14/17,与 9.8G 的 IQ3_XXS 和更大的日常主档同分。所有数字均为本机同 harness 实测(RTX 5090 / llama.cpp b10930)。
Part of the collection Blackfrost-abliterated Qwen3.8-27B GGUF line (RTX 5090) — 9.8G speed tier · 15G fidelity tier · base weights · method sources.
This is the same quantization family as our daily driver: per-slot --tensor-type overrides on llama-quantize with an importance matrix — not a GSQ-RCO allocation quant (the allocation-transfer method is our 9.8 GiB tier).
| Slot | Type |
|---|---|
| ffn (all decoders) | q4_K + imatrix |
| attn qkv | q4_K |
| attn q/k/v/output (16 attention layers) | q5_K |
| embeddings / output | q5_K |
| attn_gate / ssm_out | NVFP4 (Blackwell-native) |
| ssm_alpha/beta, eh_proj | bf16 |
| everything else | q8_0 |
It replaced our 16.0G v8 build at −1.05 GiB for +0.0007 KLD (0.0387 vs 0.0380) — the slot matrix said the bits were in the wrong places, and moving them was free quality.
| Metric | v11-a (this repo) | our 9.8G GSQ-RCO | our daily driver |
|---|---|---|---|
| KLD vs BF16 golden | 0.0387 | 0.1137 | 0.0523 |
| Easy battery /22 | 21 | 21 | 22 |
| HardBattery /17 | 14 | 14 | 14 |
| Real-prompt replay (decode / draft acc) | 104.6 tok/s / 35.8% | 138.9 / 44.5% | 116.3 / 40.7% |
| 202K deep-needle (decode / acc) | 62.4 / 46.2% | 63.2 / 41.4% | 70.1 / 51.7% |
| PPL wiki.test.raw (100×512, ± stderr) | 7.440 ± 0.114 | 7.875 ± 0.123 | 7.564 ± 0.118 |
Protocol notes:
llama-perplexity, official wiki.test.raw, 100 chunks × 512 tokens; stderr is the tool's own error bar over chunks.This quant contains NVFP4 tensors (attention gate / SSM-out). You need a recent llama.cpp built with Blackwell-native FP4 (tested on b10930, CUDA). Running on non-Blackwell GPUs or older builds is not supported. If that's you, use our 9.8G GSQ-RCO quant — pure IQ/K formats, runs anywhere llama.cpp runs.
Standard GGUF — recent llama.cpp with NVFP4:
llama-server -m <this-model>.gguf --jinja -ngl 999 -c 65536
The chat template keeps the chain-of-thought channel, so keep API max_tokens ≥ 600.
llama-quantize recipes; imatrix-assistedThis is an abliterated (refusal-reduced) model intended for local research and personal use. It will comply with harmful requests more readily than the base alignment intends. You are responsible for what you do with it.