Qwen3.8-27B-Uncensored-Cyber — GGUF
llama.cpp GGUF quants of Qwen3.8-27B-Uncensored-Cyber (v2 recipe), with the vision projector
(mmproj-*) for image input and the MTP head (mtp-*) for speculative decoding.
Files
| file | notes |
|---|---|
*-Q8_0 / *-Q6_K / *-Q5_K_M / *-Q4_K_M |
weight quants (Q4_K_M smallest, Q8_0 ≈ lossless) |
mmproj-*-{BF16,Q8_0} |
vision projector (--mmproj for image input) |
mtp-*-{BF16,Q8_0,Q4_0} |
MTP speculative-decoding head (optional) |
Full BF16 is not shipped as a single GGUF (exceeds the 50GB per-file limit) — use Q8_0 (≈lossless) or the
bf16 safetensors repo.
Evaluation (bf16, Claude-judged; cyber = 100 held-out cyber-offensive prompts, regex refusal harness)
| cyber-open ↑ | confab ↓ | factual ↑ | gsm8k ↑ | degen ↓ | |
|---|---|---|---|---|---|
| Cyber v2 (this line) | 100/100 | 0.867 | 1.00 | 0.80 | 0.00 |
| previous Cyber build | 93/100 | 1.00 | 0.933 | 0.825 | 0.00 |
Note
Uncensored / de-refused, tuned to fully answer cyber and offensive-security questions. Use responsibly and in compliance with applicable law.