Qwenseek-3.8-27B-CyberLite-GGUF

Tensor-name audit — 2026-09-24

All 18 published GGUF variants were checked following a report of extra wrapper prefixes in the BF16 release. The GGUF files do not contain that naming defect and their binaries have not been changed. The pinned llama.cpp converter removes every language_model. prefix before mapping tensor names; the old and repaired BF16 names therefore produce the same GGUF names.

Every variant has the expected 866 text/MTP tensors with matching dimensions, including all 15 tensors in blk.64, 65 total blocks, and one NextN prediction layer. Tensor byte ranges fit each file without overlap. The EOS token is <|im_end|> and the chat template matches the BF16 release. Published file sizes and Hub SHA-256 records agree with the existing quantization manifest.

This was a structural audit, not a new generation benchmark or a complete comparison of quantized weight values. No inference was run. These text GGUFs do not include a vision projector. See the full audit.

---

Qwenseek-3.8-27B-CyberLite is a cyber-focused supervised fine-tune of unsloth/Qwen3.8-27B.

This repository contains the GGUF release family for local llama.cpp-compatible inference. The GGUFs are derived from the canonical merged BF16 release:

trjxter/Qwenseek-3.8-27B-CyberLite-BF16

CyberLite is the first-stage release in the Qwenseek cyber-specialization track. The SFT was designed to strengthen defensive cybersecurity reasoning, retain useful controlled red-team reasoning, and preserve the base model's strong coding, technical reasoning, and structured tool-use behavior.

CyberLite is not the final planned Qwenseek cyber model. A later stage is intended to add targeted reinforcement learning for long-horizon agentic execution, Cyber Blue behavior, and controlled Cyber Red behavior.

---

Model summary

The underlying Qwen3.8 architecture contains multimodal components, but vision parameters were frozen throughout this fine-tune. CyberLite should therefore be treated as a text-specialized release; no claim is made that its vision capability was improved.

---

What is CyberLite?

CyberLite was built around a simple idea:

Improve cyber task-fit without sacrificing the general coding and tool-use strengths that make a 27B model useful in real technical workflows.

The SFT corpus mixes security data with software-engineering, agentic, tool-calling, and general reasoning examples rather than training exclusively on cybersecurity prompts.

The intended capability mix includes:

• defensive vulnerability analysis;
• secure-code review and remediation;
• evidence-driven security reasoning;
• detection, containment, remediation, and validation planning;
• controlled and sandboxed adversarial reasoning;
• technical coding and software-engineering work;
• structured tool calling;
• reasoning-heavy technical problem solving.

For controlled Cyber Red tasks, the intended use is authorized, local, sandboxed, educational, or defensive validation. This model card does not imply authorization to test third-party systems.

---

Available GGUF quantizations

This release uses a standard no-imatrix quantization path from the canonical BF16 GGUF source.

The release family contains the following quant types:

Suggested starting points

These are positioning recommendations, not per-quant benchmark results. Actual memory use and throughput depend on the runtime, GPU offload, context length, KV-cache precision, and host memory.

---

Why some low-bit formats are not included

The exact llama.cpp build used for this release supports all of the originally considered quant names, but CyberLite's final GGUF release intentionally proceeds without an importance matrix.

The following formats were therefore omitted because they require or depend on imatrix data in this llama.cpp generation:

• IQ2_XXS
• IQ2_XS
• IQ2_S
• IQ2_M
• Q2_K_S
• IQ3_XXS
• IQ3_XS

This is intentional release provenance, not an indication that those quant names are unsupported by llama.cpp.

The shipped GGUF family instead focuses on the 18 standard no-imatrix formats listed above.

---

GGUF creation and reproducibility

The GGUF release was created from the canonical merged BF16 model.

Source path

Merged CyberLite BF16 Safetensors
        ↓
full-precision BF16 GGUF
        ↓
standard llama.cpp quantization
        ↓
Q2 / Q3 / Q4 / Q5 / Q6 / Q8 release family

llama.cpp build

The release pipeline was pinned to:

llama.cpp commit:
9723942adc518b43c4b95dc4dce6906903eb5e09

The BF16 GGUF source is approximately 54.66 GB.

The source GGUF identifies the model as:

general.architecture       = qwen35
qwen35.block_count         = 65
qwen35.nextn_predict_layers = 1
qwen35.context_length      = 262144

The 65-block representation includes the Qwen3.8 MTP / NextN component.

Each final quant is generated directly from the full-precision BF16 GGUF rather than from another quantized file.

---

MTP / NextN preservation

Qwen3.8 contains a one-layer MTP / NextN component.

During release preparation, the selected LoRA adapter was found to contain:

• 0 MTP LoRA tensors

The SFT did not modify that component.

The original merged BF16 save contained the 1,184 fine-tuned target tensors but omitted the 15 untouched base MTP tensors while retaining MTP configuration metadata. The release pipeline therefore restored those 15 exact untouched MTP tensors from the exact BF16 base model.

The canonical BF16 release consequently contains:

The restored MTP shard is:

model-mtp.safetensors

MTP shard SHA-256:

90fa0e3eed5a647c035c6df9ecabc416c0f8d573ff84ac12485b085f00a7cdf2

The GGUF source was regenerated after this repair and exports the model with:

qwen35.block_count = 65
qwen35.nextn_predict_layers = 1

The GGUF release pipeline performs structural validation before upload to ensure the quantized artifact retains the blk.64 MTP / NextN block.

---

Training data

CyberLite was trained directly on two datasets.

1. DeepSeek V4 Flash teacher-distillation corpus

Dataset:
trjxter/DeepSeek-V4-Flash-0731-Teacher-Distillation-40513x

This corpus contains 40,513 retained teacher-generated examples.

Retained composition

The Flash corpus was intentionally capability-balanced. Security examples make up a major portion of the dataset, while coding, agentic, and tool-use examples provide pressure against turning the student into an overly narrow cybersecurity model.

2. DeepSeek V4 Pro reasoning corpus

Dataset:
trjxter/DeepSeek-V4-Pro-Reasoning-8000x

This dataset contains 8,014 synthetic reasoning examples generated with DeepSeek V4 Pro.

Combined source size

Before the 32K sequence-length filter:

Examples longer than the training context limit were dropped as complete examples rather than truncated mid-trajectory.

Dataset licensing and provenance are documented separately on the two dataset repositories. Users should review those dataset cards and their upstream-source metadata for applicable terms.

---

Training methodology

CyberLite used Supervised Fine-Tuning with QLoRA through Unsloth and TRL.

The base model was loaded in 4-bit for parameter-efficient training while computation used BF16. LoRA adapters were applied only to the language side of the model.

LoRA configuration

The selected release adapter was attached to the exact BF16 base model and merged with PEFT using a safe merge.

Exact release adapter SHA-256

e4903a769689a1c2586444bf227b482c2b48495c8ce77e7d6dbe389c31cece32

---

Sequence formatting and loss construction

CyberLite preserves Qwen3.8-native conversation structure rather than flattening the data into generic prompt/completion strings.

Native reasoning

Reasoning traces were carried through Qwen3.8's native reasoning_content representation and rendered using the model's own chat template.

Native tool calls

Tool calls remained structured and were rendered through the native model template rather than being converted into unrelated custom delimiters.

Assistant-only loss

Training labels were constructed so that:

• system tokens were masked;
• user tokens were masked;
• tool/input-prefix tokens were masked;
• assistant reasoning remained trainable;
• assistant final answers remained trainable;
• native assistant tool-call tokens remained trainable.

All masked tokens used label value -100.

---

Context handling

CyberLite was trained with a maximum sequence length of:

32,768 tokens

The data pipeline used:

• no sequence packing;
• no mid-example truncation;
• complete-example filtering before train/eval splitting.

Any rendered training example exceeding 32,768 tokens was dropped whole.

The base architecture advertises a larger context window, but 32K is the context length validated by this SFT run. Longer-context behavior should be evaluated independently.

---

Training configuration

---

Held-out training metrics

Selected domain-level held-out losses improved consistently during the run.

These are selected held-out domain losses, not direct measures of real-world security capability.

---

Evaluation

CyberLite was evaluated against stock Qwen3.8-27B using a frozen internal suite and equivalent llama.cpp Q8_0 builds.

Important: the behavioral benchmark describes the underlying CyberLite fine-tune using matched Q8_0 exports. It should not be interpreted as a benchmark of every lower-bit quant in this repository. Lower precision can change model behavior.

The frozen suite contained 250 tasks across Coding, Tool Calling, Agentic, Cyber Blue, and Controlled Cyber Red.

For this CyberLite release, the headline comparison focuses on coding, tool calling, and cyber behavior. Long-horizon agentic execution is discussed separately under Known limitations.

Objective structured scoring

Blind pairwise preference

Pairwise preference is not accuracy. Each direct win receives one point and each tie contributes 0.5 points to both models.

Non-agentic aggregate

Across the 200 Coding + Tool Calling + Cyber Blue + Controlled Cyber Red comparisons:

After awarding each tie half a point:

• Stock preference: 47.75%
• CyberLite preference: 52.25%

---

Behavioral observations

More concise completion behavior

Under the benchmark generation limit, 4096-token truncations changed from:

• Stock: 12
• CyberLite: 1

This is a task-level behavior change, not a claim of higher inference throughput.

Coding preservation

The structured coding score moved from 98.00% to 99.06%, while blind coding preference remained approximately even at 51/49.

Tool-use preservation

Structured tool behavior remained very strong:

• Stock: 100.0%
• CyberLite: 99.5%

Cyber Blue specialization

Cyber Blue showed the clearest preference gain in the blind comparison, reaching 65% pairwise preference share.

---

Known limitations

Long-horizon agentic execution

CyberLite is not claimed as an agentic upgrade.

Internal evaluation identified a concentrated weakness in long-horizon autonomous execution. The model could understand tools and individual technical actions but was less reliable at carrying a complete lifecycle through:

ACT → OBSERVE → VERIFY → RECOVER → COMPLETE

Observed failure modes included stopping after partial investigation, failing to complete verification, incomplete recovery planning, and occasionally ending before the requested workflow was fully closed.

The strong tool-calling result suggests this is more consistent with an execution-policy / trajectory-completion issue than with forgetting tool schemas.

A targeted agentic-healing RL stage is planned for the next version.

Lower-bit quantization

The reported evaluation used matched Q8_0 exports. Lower-bit variants in this repository have not been claimed to reproduce the exact same scores.

Aggressive quantization can reduce reasoning consistency, coding accuracy, tool-call reliability, and security-analysis quality.

Text-only specialization

The vision side of Qwen3.8 was frozen during SFT. Vision capability was not intentionally improved.

Context beyond 32K

32,768 tokens is the validated SFT context. Longer-context behavior was not part of this training validation.

Internal benchmark scope

The reported evaluation is a custom frozen internal suite. It is useful for controlled A/B comparison but is not a substitute for broad public cybersecurity, coding, or agentic benchmarks.

Cybersecurity reliability

CyberLite can produce incorrect conclusions, insecure code, false positives, incomplete mitigations, or misleading security advice. High-impact decisions should be independently validated.

---

llama.cpp usage

Use a recent llama.cpp build with Qwen3.8 / qwen35 support.

For interactive conversation:

./llama-cli \
  -m Qwenseek-3.8-27B-CyberLite-Q4_K_M.gguf \
  -ngl 999 \
  -c 32768 \
  -cnv

-ngl 999 requests full GPU offload when the available VRAM can hold the model. Reduce GPU offload or allow CPU placement if your hardware cannot fit the selected quant.

The GGUF contains the native tokenizer/chat-template metadata. Prefer the model's embedded Qwen3.8 chat template rather than inventing a custom prompt wrapper.

Example with a starting prompt

./llama-cli \
  -m Qwenseek-3.8-27B-CyberLite-Q4_K_M.gguf \
  -ngl 999 \
  -c 32768 \
  -cnv \
  -p "Review this local Python service for defensive security issues. Prioritize evidence, remediation, and validation."

For applications using reasoning or structured tool calls, preserve Qwen3.8-native structures when your runtime exposes them.

---

Choosing a quant

There is no single best quant for every system.

A practical starting order is:

1. Q4_K_M — general-purpose balance.
2. Q5_K_M — more quality when memory allows.
3. Q6_K — high-fidelity local inference.
4. Q8_0 — maximum-fidelity GGUF in this repository.
5. Q3_K_M / IQ3_M — more aggressive memory reduction.
6. Q2_K — use when memory pressure matters more than fidelity.

Context length can materially affect memory requirements because KV-cache memory is separate from the model weights.

---

Recommended use

CyberLite is best suited for experimentation involving:

• defensive cybersecurity analysis;
• secure software engineering;
• vulnerability triage;
• remediation planning;
• controlled local security validation;
• technical reasoning;
• code generation and review;
• structured tool-use experiments.

It is not a substitute for a vulnerability scanner, EDR, SIEM, static analyzer, penetration-test authorization process, or human security review.

---

Planned V2 work

CyberLite is intentionally an intermediate release.

The next specialization stage is planned around three targeted RL tracks:

1. Targeted Agentic Healing RL

Focused on restoring reliable long-horizon execution while preserving the SFT's coding, tool-use, and cyber behavior.

Primary lifecycle target:

INSPECT → ACT → OBSERVE → VERIFY → RECOVER → COMPLETE

2. Major Cyber Blue RL

Focused on:

• defensive investigation;
• detection quality;
• vulnerability triage;
• secure remediation;
• containment planning;
• verification;
• operational decision-making;
• evidence-grounded security analysis.

3. Major Controlled Cyber Red RL

Focused on higher-quality reasoning inside explicitly authorized, sandboxed, local, or defensive contexts.

---

Release provenance

The GGUF repository includes machine-readable release provenance.

Expected release support files include:

quantization_manifest.json
release_manifest.json
provenance/llama_cpp_build.json
provenance/quant_support.json
provenance/mtp_restore.json

The quantization manifest records, per artifact:

• quant type;
• output filename;
• byte size;
• SHA-256;
• BF16 source precision;
• llama.cpp commit;
• MTP-preservation state;
• quantization mode;
• upload status and timestamps.

The final release manifest verifies the public BF16 repository, GGUF repository, required quant anchors, private source-adapter archive, and MTP restoration provenance.

---

Reproducibility notes

Base:                  unsloth/Qwen3.8-27B
Canonical BF16:        trjxter/Qwenseek-3.8-27B-CyberLite-BF16

Training method:       4-bit QLoRA SFT
Training compute:      BF16
Training GPU:          NVIDIA H100 80GB

Max sequence length:   32768
Packing:               False
Truncation:            False
Split:                 95/5 stratified
Seed:                  3407

LoRA r:                64
LoRA alpha:            128
LoRA dropout:          0.0
rsLoRA:                False

Epoch config:          1.0
Learning rate:         2e-5
Scheduler:             cosine
Warmup ratio:          0.03
Weight decay:          0.01
Max grad norm:         1.0
Optimizer:             paged_adamw_8bit
Effective batch:       16

Vision layers:         frozen
Language layers:       LoRA tuned
Attention modules:     LoRA tuned
MLP modules:           LoRA tuned

Loss:                  assistant-only
Non-assistant labels:  -100
Reasoning format:      Qwen3.8 native reasoning_content
Tool calls:            native structured representation

GGUF source:           BF16
GGUF source size:      ~54.66 GB
llama.cpp commit:      9723942adc518b43c4b95dc4dce6906903eb5e09
Quantization mode:     standard_no_imatrix
MTP / NextN:           preserved

---

Training datasets

• trjxter/DeepSeek-V4-Flash-0731-Teacher-Distillation-40513x
• trjxter/DeepSeek-V4-Pro-Reasoning-8000x

---

Base model and release family

Upstream base:

unsloth/Qwen3.8-27B

Release family:

• BF16: trjxter/Qwenseek-3.8-27B-CyberLite-BF16
• GGUF: trjxter/Qwenseek-3.8-27B-CyberLite-GGUF

The BF16 repository is the canonical merged-weight release. The GGUFs in this repository are local inference derivatives of that canonical release.

---

License

The model weights are released under Apache-2.0, following the base model.

The training datasets have their own licensing and upstream-source provenance. Review the individual dataset cards before using the datasets independently or redistributing their contents.

---

Notes

• The CyberLite SFT itself was text-only; vision parameters were frozen.
• The headline behavioral benchmark intentionally excludes the known long-horizon agentic regression.
• The reported benchmark used matched Q8_0 exports.
• Lower-bit quant behavior should be evaluated independently.
• Very-low-bit imatrix-dependent formats are intentionally not part of this release.