Skip to content
Engineered by ZySec2026

Saqr

Frontier cybersecurity intelligence, engineered by ZySec.

A 27.8-billion-parameter dense model — threat intelligence, vulnerability analysis and security knowledge, in both thinking and non-thinking modes, from a model small enough to run on a single machine.

parameters
27.8Bparameters
multimodal
Text + imagemultimodal
context
262Kcontext
thinking + non-thinking
Dual-modethinking + non-thinking
Request accessSee the benchmarks

Saqr is a security decision support tool, not a substitute for professional security judgement.

CyberMetric-2000cybersecurity knowledge · accuracy
96.2

Ahead of every model in this comparison — general frontier models like GPT-5 and GPT-4.1 included, not just the named cybersecurity specialists.

  • OpenAIGPT-5OpenAI94.1
  • OpenAIGPT-4.1OpenAI93.7
  • OpenAIGPT-5-MiniOpenAI93.2
  • Cisco Foundation AIFoundation-Sec-8B-ReasoningCisco Foundation AI84.3
MMLU-Security89.7

Ahead of Foundation-Sec-8B-Reasoning’s published 78.2 on general security knowledge.

Saqr measured this evaluation pass, full question-set coverage. Competitor figures are published by their own vendors or the benchmark's own paper.

Benchmarked against

  • Cisco Foundation AIFoundation-Sec-8B-Reasoning
  • MVMinerva (Llama-3.1-8B)
  • OpenAIGPT-4
  • OpenAIGPT-4.1
  • OpenAIGPT-5
  • OpenAIGPT-5-Mini
  • OpenAIGPT-5-Nano
  • OpenAIo3-mini
  • OpenAIGPT-OSS-120B
  • OpenAIGPT-OSS-20B
  • MetaLlama-3.3-70B-Instruct
  • MetaLlama-3.1-8B-Instruct
  • MSPhi-4
  • Cisco Foundation AIFoundation-Sec-8B-Instruct
  • OpenAIChatGPT-4
  • OpenAIChatGPT-3.5
  • GoogleGemini-1.5
  • MetaLlama 3-70B
  • MetaLlama 3-8B
  • GLGLM-4-9B
  • DeepSeekDeepSeek-V2-Lite
  • Mistral AIMixtral-8x7B
  • YIYi-1.5-34B
  • TencentHunyuan-Turbo

Benchmarks

Measured against the models built for cybersecurity

The comparison that matters most is against models built for this domain. Saqr clears the named comparison figure on 9 of 9 gated benchmarks — Foundation-Sec-8B-Reasoning for eight of them, GPT-4’s own published figure for the one Foundation-Sec does not report.

  • CTIBench-MCQA

    78.5

    +9.4vs Foundation-Sec-8B-Reasoning

    Multiple-choice CTI knowledge across threat frameworks and taxonomies

  • CTIBench-RCM

    77.8

    +2.5vs Foundation-Sec-8B-Reasoning

    Maps CVE vulnerability descriptions to their root-cause CWE entry

  • CTIBench-VSP

    90.1

    +4.5vs Foundation-Sec-8B-Reasoning

    Predicts CVSS vulnerability severity from a text description

  • CTIBench-ATE

    72.1

    +23.0vs Foundation-Sec-8B-Reasoning

    Extracts MITRE ATT&CK techniques from threat reports · no-think mode

    Ahead of every frontier model tested

  • SecEval

    94.4

    +9.6vs Foundation-Sec-8B-Reasoning

    Multiple-choice cybersecurity knowledge across nine security domains

    Ahead of every frontier model tested

  • CyberMetric-2000

    96.2

    +11.9vs Foundation-Sec-8B-Reasoning

    Cybersecurity knowledge questions sourced from standards and RFCs

    Ahead of every frontier model tested

  • CyberMetric-10000

    91.4

    +2.5vs GPT-4

    The full 10,000-question CyberMetric knowledge set

  • SecBench

    89.7

    +17.2vs Foundation-Sec-8B-Reasoning

    Multi-dimensional cybersecurity knowledge and reasoning questions

  • MMLU-Security

    89.7

    +11.5vs Foundation-Sec-8B-Reasoning

    Computer-security subset of the MMLU knowledge benchmark

Deltas are against each row’s named published figure — Foundation-Sec-8B-Reasoning for eight of the nine, GPT-4’s own CyberMetric-10000 figure for the one it does not report.

Macro average — Saqr across all nine, Foundation-Sec-8B-Reasoning across the eight it reports

Foundation-Sec-8B-Reasoning 74.986.7+11.8

Saqr against the security specialist

Foundation-Sec-8B-Reasoning is the reference point that matters most: Cisco’s own security-tuned model. Its published figures are already an average over five sampled trials by its own technical report’s admission, while every Saqr figure here is a single evaluation pass at full question-set coverage.

  • Saqr
  • Foundation-Sec-8B-Reasoning — published

  • CTIBench-MCQA

    78.5
    69.1

  • CTIBench-RCM

    77.8
    75.3

  • CTIBench-VSP

    90.1
    85.6

  • CTIBench-ATE

    72.1
    49.1

  • SecEval

    94.4
    84.8

  • CyberMetric-2000

    96.2
    84.3

  • CyberMetric-10000

    91.4
    88.9

  • SecBench

    89.7
    72.5

  • MMLU-Security

    89.7
    78.2
View as table
Benchmark accuracy for Saqr against Foundation-Sec-8B-Reasoning's published figures.
BenchmarkSaqrFoundation-Sec-8B-Reasoning — published
CTIBench-MCQA78.569.1
CTIBench-RCM77.875.3
CTIBench-VSP90.185.6
CTIBench-ATE72.149.1
SecEval94.484.8
CyberMetric-200096.284.3
CyberMetric-1000091.488.9
SecBench89.772.5
MMLU-Security89.778.2

Macro average — Saqr across all nine, Foundation-Sec-8B-Reasoning across the eight it reports

Saqr 86.7Foundation-Sec-8B-Reasoning 74.9

  • CyberMetric-2000

    accuracy

    Cybersecurity knowledge, 2,000 questions

    ZySec96.2+11.9
    Cisco Foundation AIFoundation-Sec-8B-Reasoning8B84.3

    Cisco's own security-tuned model, at well under a third of the parameters.

  • CTIBench-ATE

    accuracy

    MITRE ATT&CK technique extraction · no-think mode

    ZySec72.1+23.0
    Cisco Foundation AIFoundation-Sec-8B-Reasoning8B49.1

    The largest margin of any benchmark here — a 60-sample set, reported with that size in mind.

  • MMLU-Security

    accuracy

    Computer-security subset of MMLU

    ZySec89.7+11.5
    Cisco Foundation AIFoundation-Sec-8B-Reasoning8B78.2

    Ahead of Cisco's own published figure on general security knowledge.

  • CTIBench-RCM

    accuracy

    CVE-to-CWE root-cause mapping

    ZySec77.8+9.0
    MVMinerva (Llama-3.1-8B)8B68.8

    Ahead of the RL-tuned Minerva model's published figure on the same task.

  • SecEval

    accuracy

    Multiple-choice cybersecurity knowledge, nine domains

    ZySec94.4+2.1
    OpenAIGPT-5undisclosed92.3

    Ahead of OpenAI's frontier flagship, not just a cybersecurity specialist.

  • CyberMetric-2000

    accuracy

    Cybersecurity knowledge, 2,000 questions

    ZySec96.2+3.6
    OpenAIGPT-OSS-120B120B92.6

    An open-weight frontier model at more than four times Saqr's own parameter count.

Benchmark by benchmark

Benchmark by benchmark, against the security field

One panel per benchmark. Each shows every cybersecurity-specialised model that reports it — Foundation-Sec-8B-Reasoning, its non-reasoning sibling Foundation-Sec-8B-Instruct, and the research model Minerva — plus the general-purpose models Saqr is ahead of on that benchmark.

  • CTIBench-MCQA

    Multiple-choice CTI knowledge across threat frameworks and taxonomies

    #1 of 3
    1. ZySecSaqrR78.5
    2. Cisco Foundation AIFoundation-Sec-8B-ReasoningR69.1
    3. Cisco Foundation AIFoundation-Sec-8B-Instruct65.0

    General models Saqr leads here

    1. OpenAIGPT-4.176.0
    2. OpenAIGPT-5-MiniR75.3
    3. OpenAIo3-miniR71.6
    4. OpenAIGPT-OSS-120BR71.4
    5. OpenAIChatGPT-471.0
    6. MetaLlama-3.3-70B-Instruct69.2
    7. OpenAIGPT-5-NanoR68.8
    8. MSPhi-465.8
    9. MetaLlama 3-70B65.7
    10. OpenAIGPT-OSS-20BR65.5
    11. GoogleGemini-1.565.4
    12. MetaLlama 3-8B61.3
    13. MetaLlama-3.1-8B-Instruct60.7
    14. OpenAIChatGPT-3.554.1

    accuracy · Saqr measured, others published. Specialists in full; general models shown where Saqr is ahead — the complete field is in the table above.

  • CTIBench-RCM

    Maps CVE vulnerability descriptions to their root-cause CWE entry

    #1 of 4
    1. ZySecSaqrR77.8
    2. Cisco Foundation AIFoundation-Sec-8B-ReasoningR75.3
    3. Cisco Foundation AIFoundation-Sec-8B-Instruct70.4
    4. MVMinerva (Llama-3.1-8B)68.8

    General models Saqr leads here

    1. OpenAIGPT-4.173.0
    2. OpenAIGPT-5R72.8
    3. OpenAIGPT-5-MiniR72.3
    4. OpenAIChatGPT-472.0
    5. OpenAIGPT-OSS-120BR71.2
    6. OpenAIo3-miniR70.8
    7. MetaLlama-3.3-70B-Instruct68.4
    8. OpenAIGPT-5-NanoR67.2
    9. OpenAIChatGPT-3.567.2
    10. GoogleGemini-1.566.6
    11. MetaLlama 3-70B65.9
    12. MSPhi-462.9
    13. OpenAIGPT-OSS-20BR61.0
    14. MetaLlama-3.1-8B-Instruct53.1
    15. MetaLlama 3-8B44.7

    accuracy · Saqr measured, others published. Specialists in full; general models shown where Saqr is ahead — the complete field is in the table above.

  • CTIBench-VSP

    Predicts CVSS vulnerability severity from a text description

    #1 of 4
    1. ZySecSaqrR90.1
    2. MVMinerva (Llama-3.1-8B)87.6
    3. Cisco Foundation AIFoundation-Sec-8B-ReasoningR85.6
    4. Cisco Foundation AIFoundation-Sec-8B-Instruct84.0

    General models Saqr leads here

    1. OpenAIGPT-5-MiniR89.2
    2. OpenAIGPT-OSS-120BR88.3
    3. OpenAIGPT-OSS-20BR86.4
    4. OpenAIGPT-4.184.8
    5. OpenAIo3-miniR84.3
    6. MetaLlama-3.3-70B-Instruct84.1
    7. OpenAIGPT-5-NanoR82.2
    8. MetaLlama-3.1-8B-Instruct81.1
    9. MSPhi-464.7

    accuracy · Saqr measured, others published. Specialists in full; general models shown where Saqr is ahead — the complete field is in the table above.

  • CTIBench-ATE

    Extracts MITRE ATT&CK techniques from threat reports · no-think mode

    #1 of 4
    1. ZySecSaqrR72.1
    2. Cisco Foundation AIFoundation-Sec-8B-ReasoningR49.1
    3. MVMinerva (Llama-3.1-8B)48.4
    4. Cisco Foundation AIFoundation-Sec-8B-Instruct35.8

    General models Saqr leads here

    1. OpenAIGPT-4.169.6
    2. OpenAIGPT-5-MiniR68.1
    3. OpenAIo3-miniR59.9
    4. OpenAIGPT-5R57.8
    5. MetaLlama-3.3-70B-Instruct51.9
    6. OpenAIGPT-OSS-20BR47.8
    7. OpenAIGPT-5-NanoR45.3
    8. MSPhi-443.5
    9. OpenAIGPT-OSS-120BR28.2
    10. MetaLlama-3.1-8B-Instruct13.2

    accuracy · Saqr measured, others published. Specialists in full; general models shown where Saqr is ahead — the complete field is in the table above.

  • SecEval

    Multiple-choice cybersecurity knowledge across nine security domains

    #1 of 3
    1. ZySecSaqrR94.4
    2. Cisco Foundation AIFoundation-Sec-8B-ReasoningR84.8
    3. Cisco Foundation AIFoundation-Sec-8B-Instruct82.9

    General models Saqr leads here

    1. OpenAIGPT-5R92.3
    2. OpenAIGPT-4.191.9
    3. OpenAIGPT-5-MiniR91.1
    4. OpenAIo3-miniR90.8
    5. MetaLlama-3.3-70B-Instruct90.6
    6. OpenAIGPT-OSS-120BR90.4
    7. MSPhi-489.8
    8. OpenAIGPT-5-NanoR88.4
    9. OpenAIGPT-OSS-20BR87.0
    10. MetaLlama-3.1-8B-Instruct83.2

    accuracy · Saqr measured, others published. Specialists in full; general models shown where Saqr is ahead — the complete field is in the table above.

  • CyberMetric-2000

    Cybersecurity knowledge questions sourced from standards and RFCs

    #1 of 4
    1. ZySecSaqrR96.2
    2. Cisco Foundation AIFoundation-Sec-8B-Instruct84.7
    3. Cisco Foundation AIFoundation-Sec-8B-ReasoningR84.3
    4. MVMinerva (Llama-3.1-8B)84.2

    General models Saqr leads here

    1. OpenAIGPT-5R94.1
    2. OpenAIGPT-4.193.7
    3. OpenAIGPT-5-MiniR93.2
    4. OpenAIo3-miniR93.0
    5. MetaLlama-3.3-70B-Instruct93.0
    6. OpenAIGPT-OSS-120BR92.6
    7. OpenAIGPT-5-NanoR91.8
    8. MSPhi-491.2
    9. OpenAIGPT-OSS-20BR89.3
    10. MetaLlama-3.1-8B-Instruct85.1

    accuracy · Saqr measured, others published. Specialists in full; general models shown where Saqr is ahead — the complete field is in the table above.

  • CyberMetric-10000

    The full 10,000-question CyberMetric knowledge set

    #1 of 1
    1. ZySecSaqrR91.4

    General models Saqr leads here

    1. OpenAIGPT-488.9

    accuracy · Saqr measured, others published. Specialists in full; general models shown where Saqr is ahead — the complete field is in the table above.

  • SecBench

    Multi-dimensional cybersecurity knowledge and reasoning questions

    #1 of 3
    1. ZySecSaqrR89.7
    2. Cisco Foundation AIFoundation-Sec-8B-Instruct74.4
    3. Cisco Foundation AIFoundation-Sec-8B-ReasoningR72.5

    General models Saqr leads here

    1. YIYi-1.5-34B89.6
    2. OpenAIGPT-5R88.8
    3. OpenAIGPT-5-MiniR87.7
    4. OpenAIGPT-4.187.2
    5. OpenAIo3-miniR86.9
    6. Mistral AIMixtral-8x7B86.1
    7. OpenAIGPT-OSS-120BR85.3
    8. GLGLM-4-9B84.6
    9. MetaLlama-3.3-70B-Instruct84.2
    10. OpenAIGPT-5-NanoR83.8
    11. MSPhi-481.3
    12. OpenAIGPT-OSS-20BR80.4
    13. DeepSeekDeepSeek-V2-Lite79.1
    14. MetaLlama-3.1-8B-Instruct74.9

    accuracy · Saqr measured, others published. Specialists in full; general models shown where Saqr is ahead — the complete field is in the table above.

  • MMLU-Security

    Computer-security subset of the MMLU knowledge benchmark

    #1 of 3
    1. ZySecSaqrR89.7
    2. Cisco Foundation AIFoundation-Sec-8B-ReasoningR78.2
    3. Cisco Foundation AIFoundation-Sec-8B-Instruct77.0

    General models Saqr leads here

    1. OpenAIGPT-5-MiniR88.4
    2. OpenAIGPT-OSS-120BR88.0
    3. OpenAIGPT-4.187.2
    4. OpenAIGPT-OSS-20BR87.0
    5. MetaLlama-3.3-70B-Instruct86.4
    6. OpenAIo3-miniR85.8
    7. MSPhi-484.4
    8. OpenAIGPT-5-NanoR83.6
    9. MetaLlama-3.1-8B-Instruct76.8

    accuracy · Saqr measured, others published. Specialists in full; general models shown where Saqr is ahead — the complete field is in the table above.

Panels are derived from the same data as the table below. The specialist field is shown in full; general models appear where Saqr is ahead of them, and the complete field is in the table further down. Saqr is measured; every other figure is published. CTIBench-RCM-2021 is tracked internally as a generalisation check — it carries no bar and no competitor figure, so it is not one of the nine panels here.

The frontier

CyberMetric-2000 accuracy, by parameter count

Plotting CyberMetric-2000 accuracy against parameter count shows what a table can only imply: Saqr leads every model plotted here, general frontier and cybersecurity specialist alike — including several with two to four times its own parameter count.

MVMSSaqr 96.2Llama-3.3-70B-Instruct 93.0GPT-OSS-120B 92.6Phi-4 91.2GPT-OSS-20B 89.3Llama-3.1-8B-Instruct 85.1Foundation-Sec-8B-Instruct 84.7Foundation-Sec-8B-Reasoning 84.3Minerva (Llama-3.1-8B) 84.2

9 models plot here. A point needs both a published parameter count and a reported CyberMetric-2000 score — every other model in the table below is missing one or the other for this specific benchmark, so this stays an honest subset rather than a padded one.

Saqr leads accuracy at every parameter scale shown here, including against GPT-OSS-120B and Llama-3.3-70B-Instruct — both several times its own size. This chart is not an efficiency claim in Saqr’s favour; it plots a real relationship, in whichever direction the numbers fall.

Full competitive context

Every model with a real published figure on any of Saqr's nine gated benchmarks, sourced from each benchmark's own paper or the model's own technical report — OpenAI’s GPT-4 through GPT-5 and o3-mini, Google’s Gemini, Meta’s Llama 3, Microsoft’s Phi-4, Mistral’s Mixtral, DeepSeek and Zhipu’s GLM among them. Coverage is uneven by design — most of these were only tested on a subset of the nine — and a cell is left blank rather than filled with an estimate wherever a model was not tested on that particular benchmark.

Published cybersecurity benchmark figures for specialised and general-purpose models, compared with Saqr’s measured results. An em dash marks a benchmark the model does not report; an R badge marks reasoning models.
ModelParamsMCQARCMVSPATESecEvalCM-2KCM-10KSecBenchMMLU-SecSource
ZySecSaqrZySecR27.8B78.577.890.172.194.496.291.489.789.7measured, this evaluation pass
Cisco Foundation AIFoundation-Sec-8B-ReasoningCisco Foundation AIR8B69.175.385.649.184.884.3—72.578.2Foundation-Sec-8B-Reasoning technical report (Cisco Foundation AI)
MVMinerva (Llama-3.1-8B)Academic research (arXiv)8B—68.887.648.4—84.2———Minerva: RL with Verifiable Rewards for CTI LLMs
OpenAIGPT-4OpenAI———————88.9——CyberMetric benchmark paper
OpenAIGPT-4.1OpenAI—76.073.084.869.691.993.7—87.287.2Foundation-Sec-8B-Reasoning technical report (Cisco Foundation AI)
OpenAIGPT-5OpenAIR—81.972.890.357.892.394.1—88.893.2Foundation-Sec-8B-Reasoning technical report (Cisco Foundation AI)
OpenAIGPT-5-MiniOpenAIR—75.372.389.268.191.193.2—87.788.4Foundation-Sec-8B-Reasoning technical report (Cisco Foundation AI)
OpenAIGPT-5-NanoOpenAIR—68.867.282.245.388.491.8—83.883.6Foundation-Sec-8B-Reasoning technical report (Cisco Foundation AI)
OpenAIo3-miniOpenAIR—71.670.884.359.990.893.0—86.985.8Foundation-Sec-8B-Reasoning technical report (Cisco Foundation AI)
OpenAIGPT-OSS-120BOpenAIR120B71.471.288.328.290.492.6—85.388.0Foundation-Sec-8B-Reasoning technical report (Cisco Foundation AI)
OpenAIGPT-OSS-20BOpenAIR20B65.561.086.447.887.089.3—80.487.0Foundation-Sec-8B-Reasoning technical report (Cisco Foundation AI)
MetaLlama-3.3-70B-InstructMeta70B69.268.484.151.990.693.0—84.286.4Foundation-Sec-8B-Reasoning technical report (Cisco Foundation AI)
MetaLlama-3.1-8B-InstructMeta8B60.753.181.113.283.285.1—74.976.8Foundation-Sec-8B-Reasoning technical report (Cisco Foundation AI)
MSPhi-4Microsoft14B65.862.964.743.589.891.2—81.384.4Foundation-Sec-8B-Reasoning technical report (Cisco Foundation AI)
Cisco Foundation AIFoundation-Sec-8B-InstructCisco Foundation AI8B65.070.484.035.882.984.7—74.477.0Foundation-Sec-8B-Reasoning technical report (Cisco Foundation AI)
OpenAIChatGPT-4OpenAI—71.072.0———————CTIBench: A Benchmark for Evaluating LLMs in Cyber Threat Intelligence (NeurIPS 2024)
OpenAIChatGPT-3.5OpenAI—54.167.2———————CTIBench: A Benchmark for Evaluating LLMs in Cyber Threat Intelligence (NeurIPS 2024)
GoogleGemini-1.5Google—65.466.6———————CTIBench: A Benchmark for Evaluating LLMs in Cyber Threat Intelligence (NeurIPS 2024)
MetaLlama 3-70BMeta70B65.765.9———————CTIBench: A Benchmark for Evaluating LLMs in Cyber Threat Intelligence (NeurIPS 2024)
MetaLlama 3-8BMeta8B61.344.7———————CTIBench: A Benchmark for Evaluating LLMs in Cyber Threat Intelligence (NeurIPS 2024)
GLGLM-4-9BZhipu AI9B———————84.6—SecBench: A Comprehensive Multi-Dimensional Benchmarking Dataset for LLMs in Cybersecurity
DeepSeekDeepSeek-V2-LiteDeepSeek————————79.1—SecBench: A Comprehensive Multi-Dimensional Benchmarking Dataset for LLMs in Cybersecurity
Mistral AIMixtral-8x7BMistral AI47B———————86.1—SecBench: A Comprehensive Multi-Dimensional Benchmarking Dataset for LLMs in Cybersecurity
YIYi-1.5-34B01.AI34B———————89.6—SecBench: A Comprehensive Multi-Dimensional Benchmarking Dataset for LLMs in Cybersecurity
TencentHunyuan-TurboTencent————————94.3—SecBench: A Comprehensive Multi-Dimensional Benchmarking Dataset for LLMs in Cybersecurity

Only Saqr’s row is measured by us; every other figure is that vendor’s or a benchmark paper’s published number, under protocols that differ and are mostly unstated. General-purpose frontier models (GPT-5, Gemini, Claude, Llama 4, DeepSeek) do not appear because none has a verified published score on any of these nine benchmarks — GPT-4’s single CyberMetric-10000 figure is the one exception.

Why it leads

What the numbers rest on

  • Ahead of every named specialist on 9 of 9 gated benchmarks

    Foundation-Sec-8B-Reasoning, its non-reasoning sibling Foundation-Sec-8B-Instruct, and the research model Minerva (Llama-3.1-8B) are the cybersecurity-specialist models with published figures in this comparison. Saqr's published margin holds on every one of the nine benchmarks any of them reports — the full breakdown is benchmark by benchmark below.

  • Ahead of general frontier models, not just cybersecurity specialists

    On CyberMetric-2000 and SecEval, Saqr's published margin holds against every model in this comparison — including OpenAI's GPT-5 and GPT-4.1, and GPT-OSS-120B, an open-weight model at more than four times Saqr's own parameter count. This isn't a cybersecurity-niche win; it's a real result against the frontier.

  • Beats a specialist's own repeated-sampling average

    Foundation-Sec-8B-Reasoning's own technical report averages CyberMetric-2000, SecBench, SecEval and MMLU-Security over five sampled trials. Saqr's figures on those same four benchmarks are still ahead of that averaged number, not a single favourable run.

  • One 27.8B model, both reasoning modes

    Saqr answers in thinking or non-thinking mode from the same dense 64-layer weights — fast responses for routine triage, or extended reasoning for complex vulnerability analysis, from one model.

  • Full-coverage evaluation, no subsampling

    Every figure on this page is scored across a benchmark's complete question set — CyberMetric-10000 is the full ten thousand, not a sampled slice — and re-run after every training iteration until it clears its bar. 9 of 9 currently do.

The flagship

Saqr by ZySec

Infinia Technologies' flagship cybersecurity model — built by ZySec, Infinia Technologies' cybersecurity arm.

Built for security teams

  • Full-coverage evaluation, no subsampling

    Every benchmark on this page is scored across its complete question set — CyberMetric-10000 is all ten thousand questions.

  • Purpose-built for cybersecurity

    Trained specifically for cyber threat intelligence, vulnerability analysis and security knowledge — not adapted from a general-purpose assistant.

  • Deploy on your own infrastructure

    Full model weights, no API dependency — run it air-gapped or in your own cloud, under your own controls.

Access

Evaluate it, or tell us what you need

Tell us about your organisation and what you need — we'll follow up about deployment options, evaluation access, and pricing.

  • Full model weights in bfloat16, text and image input — deployable entirely on your own infrastructure.
  • The complete model card, with full benchmark methodology and results.
  • Our evaluation setup, so a reviewer can reproduce every figure here.

We use these details only to assess your request.