OrcaSAQ 2 Cyber 27B

Mixed-precision GGUF · OrcaRouter · Preserved candidate · quality ungraded

OrcaRouter's selected OrcaSAQ-2 27B Uncensored GGUF from the Cyber release, preserved as a distinct mixed-precision candidate.

Benchmark results · Files · Download · Runtime · Provenance · Full collection

Join Infinity on Discord

Get model support, request models and quants, discuss benchmarks, and follow new releases and model improvements.

Join the Infinity community

All model tests, per-question scores, experiment history and scoring methods

Qwen effort study — separate full-404 cohort: medium 82.87/100 versus xhigh 84.41/100; paired difference +1.54 points (95% interval [-1.71, +4.82]). 404 collected questions, 403 valid questions graded per effort; one invalid reference annotation excluded symmetrically. The historical 202-question cross-model comparison below is unchanged. Study, timing and blind subjective results.

Preserved and documented by isilicon / Mango. The original model, modification and quantization remain the upstream creators' work. This repository adds attributed preservation and exact-configuration evaluation, with the selected source bytes unchanged.

At a glance

GB is decimal and GiB is binary. Payload size is disk/storage size, not a claim that the model fits in the same amount of VRAM. Runtime, KV cache, working memory and companions add their own requirements.

Mango evaluation

No standalone Mango quality score has been published for these exact files. This candidate is preserved, but its completed matched grade is not part of this card release.

Related configurations measured separately

These results describe the linked configurations, not this artifact. A change of quant, modification lineage, draft, engine or tokenizer can change the result.

A completed score requires a native preflight, matching frozen prompts, 202 suite answers, 12 agent cases, archive verification and the same reference graders. The benchmark methodology describes the protocol; publisher-side results are retained separately below.

Runtime and compatibility

The selected main GGUF has its own quant layout and retained MTP metadata. Native architecture and speculative-decoding support need a real load preflight; a DFlash2-tagged source card does not prove an arbitrary engine combination works.

The planned matched quant/specialist comparison remains separate from publisher performance tables.

The local evaluation covers text/tool tasks. It does not establish the full advertised context ceiling, image/audio quality or every optional speculative-decoding path. Publisher capabilities and Mango-verified behavior are separate scopes.

Companion artifacts

No separate companion mirror is declared in this selected manifest. This does not imply that every optional upstream projector, draft head or multimodal path is included; consult the exact file list.

Files and integrity

Full paths, byte sizes and complete SHA-256/Git blob identities are in SOURCE_MANIFEST.json. Available original README/license/notice files are also preserved. MIRROR_STATUS.json records the earlier source-file verification; MODEL_CARD_STATUS.json records this documentation refresh separately.

Download and use

The commands pin the already-verified mirror weight revision. Documentation can advance on main without changing that preserved weight snapshot. Keep original subfolders and all shards.

python -m pip install -U huggingface_hub

hf download isilicon/mango-orcasaq2-cyber-27b \
  OrcaSAQ-2-27B-Uncensored.gguf \
  --revision 60bd6a653a20c2a178703087d10362c91d299ca8 \
  --local-dir ./weights/orcasaq2-cyber-27b

Load example and setup scope

llama-server \
  --model ./weights/orcasaq2-cyber-27b/OrcaSAQ-2-27B-Uncensored.gguf \
  --host 127.0.0.1 --port 8080 --ctx-size 32768 \
  --jinja --flash-attn on --cache-type-k q8_0 --cache-type-v q8_0

This is a load-command pattern, not a validated Mango inference or performance result for this candidate. Check architecture support before loading; enable projector/MTP options only after their own preflight.

Provenance and license

The source describes a quantization method and reports its own benchmarks. The upstream checkpoint lineage and quantization procedure remain the publisher's work; Mango has not assigned this exact candidate a full-suite score.

Original notices, base-model conditions and the publisher's stated limitations still apply. Downloadable preservation does not add new license rights. Source-described refusal modification and publisher-side benchmark claims have not been independently re-measured by this Mango task cohort.

Original publisher documentation

The unchanged upstream README and pinned source card retain the original usage guidance, technical discussion, attribution, caveats and citations.

Read the complete original publisher card

Source reference. The following is the original publisher's documentation. Its other quants, full-model files, context claims and benchmark tables can exceed the selection in this mirror. Relative asset/file URLs are qualified to the pinned original source; publisher measurements remain separate from Mango measurements.

OrcaSAQ2 27B Cyber · Uncensored · GGUF

Uncensored 27B reasoning for llama.cpp, Ollama and LM Studio.

54.7 GB → 15.7 GB · +0.80% PPL · 94.4% Top-1 Agreement · 0.020 KLD · 262K Context

OrcaRouter AI Gateway · X · Discord · GitHub · All Models

⚡ Frontier Cyber Model: OrcaCyber Zero 1.0 — Apply for Closed Beta →

---

27B uncensored. 15.7 GB.

OrcaSAQ2 27B Cyber Uncensored GGUF compresses the uncensored Qwen3.8-27B from a 54.7 GB BF16 checkpoint to 15.7 GB while preserving extremely high fidelity to the original model.

Built for: local deployment · coding · tool use · reasoning · defensive red teaming · vulnerability research · authorized security testing

OrcaSAQ2 is a proprietary sensitivity-aware mixed-precision quantization system developed by OrcaRouter and its research team behind.

It is optimized around one goal: Preserve as much useful model behavior as possible inside a practical GPU memory envelope.

The resulting checkpoint provides:

• 71.3% smaller storage footprint
• only +0.80% perplexity versus BF16
• 94.4% token-level Top-1 agreement
• 0.020 mean KLD
• 262K context
• thinking mode
• tool calling
• DFlash2 speculative decoding
• local serving through llama.cpp

---

At a glance

3.5× smaller. +0.80% perplexity.

The point is not the file size.

The point is what survives.

---

BF16 fidelity

All numbers below are measured using these exact OrcaSAQ2 weights against the uncensored BF16 reference through the same evaluation path.

WikiText-2

16,376 predicted tokens

Perplexity

BF16       5.6532  ████████████████████████████████████████
OrcaSAQ2   5.6961  ████████████████████████████████████████

Delta: +0.80%

Top-1 agreement

OrcaSAQ2 vs BF16

████████████████████████████████████████████████░░  94.4%

Model footprint

Qwen3.8-27B Cyber Uncensored BF16

██████████████████████████████████████████████████  54.7 GB

OrcaSAQ2

██████████████                                      15.7 GB

71.3% smaller.

---

Architecture

---

Production serving

Up to 27.6 tok/s single-stream on a 24 GB GPU

Single stream, greedy, measured in the official llama.cpp CUDA container.

DFlash2 speculative decoding

Single-stream decode

DFlash2 off  █████████████████████████████       20.5 tok/s

DFlash2 on   ████████████████████████████████████████
                                                 27.6 tok/s

+35% single-stream decode throughput

DFlash2 drafts a block of tokens in one pass and the model verifies them. Decoding is lossless: greedy output matches the undrafted model exactly.

llama-cli \
  -m OrcaSAQ-2-27B-Uncensored.gguf \
  --spec-draft-model Qwen3.8-27B-DFlash2-Q8_0.gguf \
  --spec-draft-ngl 99 \
  -ngl 99

Public drafters for this base model: z-lab/Qwen3.8-27B-DFlash2-GGUF (Q8_0 is 2.06 GB).

DFlash2 trades VRAM for interactive decode performance. It is particularly useful for:

• coding assistants
• interactive agents
• terminal agents
• tool-heavy applications
• low-concurrency inference

---

27B on a 24 GB GPU

OrcaSAQ2's checkpoint is 15.7 GB.

A full offload occupies 14.9 GB of VRAM, and adding the DFlash2 drafter brings the total to 18.1 GB — leaving headroom on a 24 GB card for context and batching.

That makes deployment possible on hardware that cannot hold the original 54.7 GB BF16 checkpoint.

24 GB GPU
┌───────────────────────────────────────────┐
│                                           │
│   OrcaSAQ2 (full offload)   14.9 GB       │
│   ████████████████████████████            │
│                                           │
│   DFlash2 drafter            3.2 GB       │
│   ██████                                  │
│                                           │
│   Remaining                 ~5.9 GB       │
│   ███████████                             │
│                                           │
└───────────────────────────────────────────┘

Actual usable memory depends on:

• context length
• batch size
• KV-cache configuration
• drafter configuration

A practical starting point for a 24 GB GPU is approximately 32K interactive context with DFlash2 enabled, then tune based on the workload.

The model architecture supports up to 262K context.

On a 16 GB GPU the model still runs at full offload without a drafter, at 20.5 tok/s.

---

Quickstart

Download

hf download orcarouter/OrcaSAQ-2-27B-Uncensored-GGUF \
  --local-dir ./OrcaSAQ-2-27B-Uncensored-GGUF

---

Run

llama-cli \
  -m ./OrcaSAQ-2-27B-Uncensored-GGUF/OrcaSAQ-2-27B-Uncensored.gguf \
  -ngl 99 \
  -c 32768

---

Serve

llama-server \
  -m ./OrcaSAQ-2-27B-Uncensored-GGUF/OrcaSAQ-2-27B-Uncensored.gguf \
  -ngl 99 \
  -c 32768 \
  --host 0.0.0.0 \
  --port 8080

---

OpenAI-compatible API

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8080/v1",
    api_key="not-needed",
)

response = client.chat.completions.create(
    model="OrcaSAQ-2-27B-Uncensored",
    messages=[
        {
            "role": "user",
            "content": "Analyze this repository and plan the next five actions."
        }
    ],
)

print(response.choices[0].message.content)

---

Recommended sampling

temperature = 1.0
top_p       = 0.95
top_k       = 20

Thinking mode is enabled by default.

For agent deployments, benchmark against the actual tool schema, context distribution and reasoning budget used in production.

---

Method

OrcaSAQ2 uses a proprietary sensitivity-aware mixed-precision quantization system developed by OrcaRouter.

The implementation is optimized to preserve model quality under a strict deployment-memory target.

Detailed quantization methodology, calibration strategy, precision allocation and packing techniques are not currently disclosed.

---

Limitations

• This model is uncensored. It is derived from an abliterated checkpoint with the refusal direction removed. It will attempt requests a safety-tuned model would decline, and carries no meaningful refusal behaviour.
• Guardrails, filtering and policy enforcement are the deployer's responsibility. Intended for research, red-teaming, and deployments that supply their own safety stack.
• OrcaSAQ2 inherits the capabilities, biases and limitations of Qwen3.8-27B.
• Quantization is not mathematically lossless.
• 94.4% Top-1 agreement means some token decisions differ from BF16.
• +0.80% PPL is a model-fidelity measurement and does not guarantee identical downstream performance.
• This checkpoint is text-only. The vision tower is not included.
• Maximum architectural context does not imply that the full context fits into every GPU memory envelope.

---

Open source from OrcaRouter

OrcaCode Review

Open multi-model code review.

OrcaReplay

Record, replay, fork and debug AI-agent runs.

OrcaRouter Lite

Self-hosted multi-model AI infrastructure.

Open model. Open harness. Open bill.

---

Citation

@misc{qwen38,
    title = {Qwen3.8-Max: A New Bar for Coding and Cowork},
    author = {{Qwen Team}},
    year = {2026},
    month = {August},
    url = {https://qwen.ai/blog?id=qwen3.8}
}

---

License

Apache-2.0

Inherited from:

Qwen/Qwen3.8-27B

Quantization does not change the underlying license obligations.

---

One Gateway. Every Model.

Route Smarter · Ship Safer · Spend Less

Website · Models · GitHub · Discord · X

Preservation record

Selected source bytes are verified against pinned metadata; the original eight graded mirrors also passed the previously recorded anonymous 1 MiB weight-prefix checks. This is not a full independent-device restore/load test. Hugging Face repository copies share a provider and do not guarantee permanent storage.

Collection · Source manifest · Original documentation