license: apache-2.0
base_model: Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16
base_model_relation: quantized
library_name: gguf
tags:
- gguf
- llama.cpp
- quantized
- mixed-quant
- abliterated
- uncensored
- 27b
- qwen3.8
- nvfp4
Qwen3.8-27B Blackfrost-Abliterated · Slot-Tuned Mixed 4/5-bit GGUF (~14.95 GiB)
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.
Recipe (slot-tuned, not allocation-table)
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.
Measured (RTX 5090 / llama.cpp b10930 / same harness for every cell)
| 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:
- KLD is against saved BF16 golden logits (f16 KV, identical corpus/chunking for all quants) — a weight-fidelity indicator, not a quality ranking.
- Easy/HardBattery: knowledge quiz + open-ended generation + IFEval-style instruction following; and math with brute-force-verified ground truth + sandbox-executed coding + hard logic. HardBattery is the only battery that ever separated model generations for us — quantization recipes have never done so at these sizes.
- Replay: 30 real user prompts, paired across quants, temp 0, with the same BF16 speculative draft model.
- PPL:
llama-perplexity, officialwiki.test.raw, 100 chunks × 512 tokens; stderr is the tool's own error bar over chunks.
⚠ Requires an NVFP4-capable llama.cpp
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.
Honest limits
- The 450-case refusal-bench protocol has not been run on this specific quant (our 9.8G card has that data, including the Blackfrost official 17.8% anchor).
- Speed/draft numbers assume our BF16 DFlash speculative draft; acceptance varies by workload (code peaks, prose dips).
- 202K long-context is where our daily driver stays ahead (70.1 vs 62.4 tok/s needle decode) — pick by workload, not just size.
Usage
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.
Credits & license
- Base weights: Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16
- Quantization method: slot-tuned mixed
llama-quantizerecipes; imatrix-assisted - License inherits the base model (Apache-2.0 per Qwen lineage)
Disclaimer
This 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.