117 4 days ago

vision tools thinking
ollama run orcarouter/OrcaCyber-Nano-1.0:mix-13g

Details

4 days ago

ce21f79c8071 · 14GB

qwen35
·
27.3B
·
IQ3_S
clip
·
461M
·
F16
{ "draft_num_predict": 4, "min_p": 0, "presence_penalty": 0, "repeat_penalty": 1,

Readme

OrcaRouter OrcaCyber Nano 1.0 27B cybersecurity defense, on your laptop — reverse engineering · vulnerability analysis · secure code review · vision · 262K context

Qwen license · 2 tags · 13.9 / 16.6 GB · 16 GB or 24 GB GPU · vision bundled · thinking · MTP on

Website · Model Catalog · Hugging Face · GitHub · Discord · X


🛡️ New from Orca — OrcaCyber Zero 1.5

Post-trained for cybersecurity: vulnerability analysis, vulnerability reproduction, red teaming & offensive-security research. 98.01% pass@1 on CyberGym Level 1 · 1M context · function calling · OpenAI-compatible API.

Try it on OrcaRouter →

🔌 New from Orca — Orca Connect

Running this in Ollama on your own machine? Orca Connect puts the rig under your desk behind your usual OrcaRouter API key, next to 200+ hosted models, reachable from a laptop, a phone or a cloud agent. Outbound only: no port forwarding, no tunnel, no firewall rule. Requests your own hardware serves cost $0, and orca-connect claude points Claude Code at your rig for one session.

Connect your rig → · ▶ Watch the demo


A security engineer’s model. Runs locally.

OrcaCyber Nano 1.0 is a 27B vision-language reasoning model built for defensive security work — and packed for Ollama so it runs on your own hardware, offline, where the binaries and logs you are analyzing never leave the machine.

Built for: reverse engineering · vulnerability analysis · secure code review · triage · incident response


Why run it locally

Security work is exactly the kind of work you do not want to send to a cloud endpoint: proprietary binaries, customer logs, internal source, unreleased vulnerabilities. With Ollama the model runs on the same box as your tools, air-gapped if you need it.

  • Reverse engineering & decompilation — paste disassembly or decompiler output; recover what a stripped function does, name variables, explain control flow.
  • Vulnerability analysis — reason about exploitability, find the unsafe path, propose a patch.
  • Secure code review — audit diffs and files for injection, memory-safety and logic bugs.
  • Triage — read logs, alerts and packet captures; separate signal from noise.
  • Vision — feed screenshots, Ghidra/IDA captures, network diagrams.

It identifies as OrcaCyber Nano 1.0 AI Defense Model and keeps full general-purpose reasoning, so it is also a strong local coding and reasoning assistant.


Pick a tag

Tag Build Size Fits
mix-15.7g · latest Mix-15.7G + vision projector 16.6 GB 24 GB+ card, best quality
mix-13g Mix-13G + vision projector 13.9 GB 16 GB card

Both tags carry the built-in MTP head and the vision projector.


Usage (Ollama)

Requires Ollama v0.40.16 or later.

ollama run orcarouter/OrcaCyber-Nano-1.0             # Mix-15.7G
ollama run orcarouter/OrcaCyber-Nano-1.0:mix-13g     # 16 GB card

cURL

curl http://localhost:11434/api/chat -d '{
  "model": "orcarouter/OrcaCyber-Nano-1.0",
  "messages": [{"role": "user", "content": "Hello!"}]
}'

Python

from ollama import chat
resp = chat(
    model="orcarouter/OrcaCyber-Nano-1.0",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(resp["message"]["content"])

Image input

ollama run orcarouter/OrcaCyber-Nano-1.0 "What is in this image? ./photo.png"
  • Images: drag one into ollama run, or send it in images through the API.
  • Thinking is on by default. Ollama returns the reasoning in message.thinking. Set "think" to false, "low", "medium" or "xhigh" to change it.
  • MTP is on. The built-in MTP head drafts 4 tokens a step (draft_num_predict 4), about 3× faster decoding than without it on our H200. The DFlash2 drafter is for llama.cpp only.
  • Sampling is set to the recommendation: temperature 1.0, top_p 0.95, top_k 20.

Evaluation

Benchmark Score
MATH-500 98.60%
GSM8K 97.50%
AIME 2025 96.23%
HumanEval+ 93.29%
LiveCodeBench 92.23%
GPQA Diamond 88.28%
MBPP+ 84.66%
IFEval 82.81%
SWE-bench 78.4%
Terminal-Bench 2.1 72.5%
IFBench 56.33%
BigCodeBench 49.82%
MATH-500            █████████████████████████████████████████████████░  98.60%
GSM8K               █████████████████████████████████████████████████░  97.50%
AIME 2025           ████████████████████████████████████████████████░░  96.23%
HumanEval+          ███████████████████████████████████████████████░░░  93.29%
LiveCodeBench       ██████████████████████████████████████████████░░░░  92.23%
GPQA Diamond        ████████████████████████████████████████████░░░░░░  88.28%
MBPP+               ██████████████████████████████████████████░░░░░░░░  84.66%
IFEval              █████████████████████████████████████████░░░░░░░░░  82.81%
SWE-bench           ███████████████████████████████████████░░░░░░░░░░░  78.4%
Terminal-Bench 2.1  ████████████████████████████████████░░░░░░░░░░░░░░  72.5%
IFBench             ████████████████████████████░░░░░░░░░░░░░░░░░░░░░░  56.33%
BigCodeBench        █████████████████████████░░░░░░░░░░░░░░░░░░░░░░░░░  49.82%

Benchmark harness, settings and full result tables are being finalized and will be expanded.


Model details

Architecture Qwen3_5ForConditionalGeneration, VL hybrid
Layers / hidden 64 / 5120 — 48 linear-attention + 16 full-attention
Context 262,144 tokens
Modalities text + image/video in, text out
Speculative decoding MTP head (built in, on in Ollama); DFlash2 draft for llama.cpp only

License

The Qwen license, inherited from the base model Qwen/Qwen3.8-27B.


Elsewhere


One Gateway. Every Model.

Route Smarter · Ship Safer · Spend Less

Website · Models · GitHub · Discord · X