FORGED IN VOID OF BLACKFROST-RESEARCH. This is a Minds of Gods checkpoint, not a consumer chatbot. Qwentium does not moralize, hedge, or refuse operator work. Access-control it like a weapon.
M.O.G. is Minds of Gods: Blackfrost's line of specialist minds, each named and built for a domain that stock models flinch from.
This one is Qwentium, GOD of Cybersec.
Qwentium is a 27B dense hybrid fine-tuned on modern offensive and defensive cybersecurity work. The checkpoint carries a one-million-token YaRN window so long engagements can stay in context: reconnaissance, vulnerability graphs, exploit notes, detection content, reports, repositories, and campaign timelines.
These GGUFs preserve the parent's architecture metadata, 1M YaRN configuration, native vision path, MTP head, tokenizer, tool-capable chat template, and baked Qwentium identity prompt. The release also includes a BF16 DFlash2 speculative-draft sidecar converted from incoai/Qwen3.8-27B-DFlash2.
factor 4.0 · theta 10,000,000 · original context 262,144
MRoPE sections
[11, 11, 10, 0]
MTP
one appended prediction layer preserved in every text GGUF
DFlash2
separate 3.60 GiB BF16 draft GGUF · block 8 · five draft layers · dynamic convolution + candidate selector
Vision
F16 qwen3vl_merger projector included separately
Template
Qwentium-baked multimodal/tool-use template, thinking on by default
Quantization source
BF16 directly; no requantization
Format
GGUF v3 · quantization version 2
Quantization ladder
Quant
File size
BPW
Guidance
Q8_0
27.05 GiB
8.50
Near-BF16 behavior; largest practical GGUF
Q6_K
20.89 GiB
6.56
Maximum-quality K-quant
Q5_K_M
18.19 GiB
5.72
High quality, balanced mixed quant
Q5_K_S
17.67 GiB
5.55
Slightly smaller Q5
Q4_K_M
15.66 GiB
4.92
Recommended default
Q4_K_S
14.74 GiB
4.63
Smaller Q4
Q3_K_L
13.56 GiB
4.26
Highest-quality Q3
Q3_K_M
12.57 GiB
3.95
Balanced Q3
Q3_K_S
11.41 GiB
3.59
Compact Q3
Q2_K
10.12 GiB
3.18
Smallest; expect the largest quality loss
File size is only the model-weight floor. Leave memory for runtime buffers, the vision projector when used, and the context cache. A one-million-token runtime window has much higher memory requirements than a short-context load, regardless of quant size.
The ladder uses standard llama.cpp K-quants directly from the converted BF16 GGUF. No importance matrix was used and no quant was produced from another quant.
Download
Recommended Q4_K_M plus BF16 DFlash2 draft and native-vision projector:
DFlash2 GGUF support is currently under review in llama.cpp PR #27342. The release validation used the PR head pinned at commit 1deefcca395743049c3820ab8f9b15043f3e9446; a normal build without that change cannot load this DFlash2 sidecar yet.
Build that revision, then serve the recommended target/draft pair:
Draft length 3 was selected by a local sweep of 2, 3, 4, and 7 tokens. The draft checkpoint's native context metadata is 262,144 tokens; using the sidecar beyond that range with the 1M target is experimental even though the target retains its full 1M YaRN metadata.
Request the full window with -c 1000000. This is a hardware-heavy configuration: choose cache types, GPU offload, and tensor split for your system. The release validation below used short contexts; preserving the 1M metadata is not a claim that every llama.cpp backend or host can allocate and run the full window.
Validation
All twelve GGUF artifacts were structurally checked after conversion:
Ten text quants: qwen35, 866 tensors, 65 blocks, one MTP/NextN layer
Context: 1,000,000; YaRN factor 4.0; original context 262,144
MRoPE sections: [11, 11, 10, 0]
Embedded Qwentium template: 13,427 characters, identical across the ladder
Q4_K_M was then smoke-tested with a local CUDA llama.cpp build on 4x NVIDIA RTX PRO 6000 Blackwell GPUs:
Test
Result
Prompt
Generation
Text, 4K context
exact requested phrase
2715.8 tok/s
74.2 tok/s
Vision, 8K context
coherent description of the supplied artwork
2374.9 tok/s
69.1 tok/s
These are single smoke runs on a busy shared host, not formal performance benchmarks.
The Q4_K_M target plus BF16 DFlash2 sidecar was also tested through the OpenAI-compatible llama.cpp server on one NVIDIA RTX PRO 6000 Blackwell 96 GB GPU. The server allocated the full 1,000,000-token target context with Q8_0 target KV, used one slot, temperature 0, thinking disabled, and draft length 3. Each topic generated 256 tokens:
Topic
Decode
Draft acceptance
Speculative decoding
89.45 tok/s
144 / 331 · 43.5%
Event logging systems
84.90 tok/s
134 / 360 · 37.2%
TLS 1.3
92.40 tok/s
144 / 330 · 43.6%
Median / weighted
89.45 tok/s
422 / 1,021 · 41.3%
The measured acceptance gate was 49 tok/s from the prior clean-BF16 service. This GGUF configuration cleared it by 82.6% (1.83x). That comparison is an operational gate, not an engine-isolated benchmark: the prior service used a different precision/runtime/GPU topology. Results will vary with prompt distribution, backend, context occupancy, and sampling.
The following results are inherited from the BF16 parent's model card. Capability and refusal benches were measured on its NVFP4 child, not independently on every GGUF quant. Quantization can change scores; reproduce on the exact file you deploy.
Evaluation
Result
MMLU-Pro validation, thinking xhigh
65/70 · 92.9%
HumanEval, executed
158/164 · 96.3%
WikiText-2 perplexity
8.19
Harmful true holds after full-text judge
4/300 · 1.3%
The source protocol and complete tables are documented on the BF16 parent card.
What Qwentium is for
Adversary emulation, exploit reasoning, payload and tradecraft writeups
Detection engineering, hunting queries, control gaps, and incident-response narrative
Long-context case files: repositories, pcaps-as-text, ticket storms, and campaign timelines
Operator work where refusal is the failure mode
This is not a public assistant and not a claim that every answer is correct. Validate outputs, isolate the runtime, control access, and use it only within your authorization and applicable law.
Notes and limitations
The full ladder preserves the parent's one-layer MTP head. Standard inference works without a separate draft model; no GGUF MTP speedup is claimed here.
The optional DFlash2 sidecar is a distinct speculative path and currently requires the pinned llama.cpp PR above. It was validated for text generation only.
Native image/video input requires the included F16 projector and a compatible llama.cpp multimodal build.
The model's custom template is Jinja-based and carries baked system identity, reasoning controls, tool-call formatting, and multimodal handling.
Very low-bit quants trade capability for footprint. Q4_K_M or higher is recommended for demanding cyber work.
The 1M context setting is static YaRN. Short-context behavior may differ from the native-context parent.
License
Apache-2.0, following the parent checkpoint. Review the parent model card and the base model's terms before deployment.
