Jue & Jue-Flash jué · [tɕɥɛ˧˥] · /v1/systemone
Hanzi 决 (Trad. 決)
Mandarin Pinyin jué
IPA [tɕɥɛ˧˥]

“决” (jué, [tɕɥɛ˧˥] — high-rising 2nd tone): originally the breaching of a dike to release a decisive torrent; in classical Chinese strategy and jurisprudence, to adjudicate, resolve, and render an irrevocable judgment without hesitation.

Calibrated Cybersecurity Decision Models

A probability for every answer
to a security question.

Jue (决) and Jue-Flash (决·闪) are open-weight decision models for cybersecurity: SOC alert triage, detection engineering, threat intelligence, vulnerability management, cloud and identity security, and guardrails for AI agents. You send an alert, a log, a policy or a CVE together with a few typed questions (Choice, Score, yes/no Noul), and get back a calibrated probability for every option, in one forward pass and without generated text. Both speak the Jev /v1/systemone protocol. This page shows what they get right and what they get wrong.

Apache-2.0 2.5B Decoder

Jue (决)

MiniCPM5-2B + LoRA r32 for SOC, CTI, vulnerability, cloud and DevSecOps decisions. SecQA .973; CyberMetric-10000 .735; BFCL-v4 irrelevance .863.

LFM Open License 350M Encoder

Jue-Flash (决·闪)

LiquidAI/LFM2.5-Encoder-350M + a decision head for fast first-pass triage, about 0.1 s per request on Apple silicon. SecAlertBench .637; VulDetectBench task 2 .886.

Download the models ↓ Technical report Paired results
Known Limits

What Works, and What Does Not

Every number on this page comes from the released checkpoints. The technical report has the full account.

JUE · RELEASE CRITERIA

Three of eight criteria met

The last training step improved tool-use decisions (BFCL-v4 irrelevance .742 → .863, When2Call .775 → .836) but met only three of the eight criteria fixed before it ran. On open-set CWE questions whose candidates contain the answer, Jue now says "none of these" 28% of the time (was 4.3%). SDDL access checks and prerequisite API calls got worse, and it is under-confident on unfamiliar benchmarks by .057 to .068.

JUE · RULES AND CODE

Rule execution and repository-scale code

Yes/no questions that need a rule evaluated completely (reachability, privilege, dependency) stay well behind choice questions on the same states, and a probe found nothing extra in earlier layers. The original code goals, repository-level walking and deep logic-bug finding, were not reached.

JUE · ALERT TRIAGE

Too quick to call an alert an attack

On SecAlertBench Jue calls 354 of 527 benign alerts attacks and scores below chance. Fit a threshold on your own alerts before using it as a filter.

JUE-FLASH · SCOPE

Fast, but narrow

On the same items of 131 external tasks Jue-Flash beats Jue on 14. It truncates states longer than 4,096 tokens, fails questions that need several options at once, seldom decides that no tool should be called, and is much weaker in Chinese. Use it to pre-filter, not to decide.

BOTH · READING THE TABLES

Paired, on the same questions

Both models are scored on exactly the same items with each benchmark's own metric. Where Jue was read on a 750-question sample, both columns use that sample. Answers are the argmax over options, so some numbers are not comparable to published leaderboards; the notes say which.

Two Models, One Protocol

Jue (2.5B Decoder) and Jue-Flash (350M Encoder)

Both models speak the same wire protocol (POST /v1/systemone). Run Jue-Flash first and pass uncertain cases to Jue, or use Jue alone; the two-tier setup has not been measured end to end.

◆ 2.5B Causal Decoder · Open weights

Jue 决 · 重剑无锋

MiniCPM5-2B with a LoRA adapter; a shared state prefix plus one branch per question, answered from the language-model head's rows of the option letters ("letter" readout).

  • Backbone openbmb/MiniCPM5-2B (LoRA r = 32, α = 64)
  • Context & Head Requests up to 8,192 tokens in training, never truncated; letter readout
  • Core Strength Security policy and configuration decisions, CTI and vulnerability questions, tool-call decisions; weak on complete rule execution
  • Calibration One global temperature (T 1.091); well calibrated on familiar formats, under-confident on unfamiliar benchmarks
.973 SecQA v1
.863 BFCL-v4 irrelevance
.735 CyberMetric-10000
⚡ 350M Bidirectional Encoder

Jue-Flash 决·闪 · 疾电破阵

Hybrid convolution + bidirectional attention encoder (LiquidAI/LFM2.5-Encoder-350M) with a dual-path decision head (pooled bilinear + token MaxSim) and an ordinal coupler for Score questions.

  • Backbone LiquidAI/LFM2.5-Encoder-350M (16-Layer Hybrid LivConv + Bi-Attn)
  • Context & Head DualPathDecisionHead (4-Depth Gate + 2-Hop Cross-Attn + MaxSim)
  • Inference Speed About 0.1 s per request on Apple silicon after warm-up; states truncated at 4,096 tokens
  • Core Strength Fast first-pass triage; behind Jue on most external benchmarks (see the paired tables)
  • Calibration One temperature per question type plus a yes/no bias
.637 SecAlertBench
.886 VulDetectBench T2
.697 CTIBench RCM
Mathematical Foundations

Three Typed Primitives of System One Judgment

Every cybersecurity triage, code audit, or policy compliance check can be decomposed into three canonical decision primitives evaluated simultaneously over a shared state $S$ in a single forward pass.

1. Choice (K-Way) 择一 · 互斥多选

Selects one option among $K \in [2, 255]$ dynamically supplied candidates (e.g., CWE root cause, ATT&CK technique, routing target, or policy clause).

$$\begin{aligned} P(y = k \mid S, q) &= \text{softmax}_k\!\left(\frac{z_k(S, q, c_k)}{T_{\text{choice}}}\right) \\[3pt] \text{conf}(P) &= 1 - \frac{H(P)}{\log K} \end{aligned}$$

Jue: the dot product of the final hidden state with the option letters' rows of the language-model head. Jue-Flash: pooled bilinear scores combined with token-level MaxSim.

2. Score (Ordinal) 品第 · 序数评定

Evaluates ordered severity or rubric scales of length $L \in [2, 10]$ (e.g., CVSS impact, SSVC remediation priority, alert severity, or relevance grade).

$$\begin{aligned} z_k^{\text{ord}} &= z_k^{\text{span}} - e^{\gamma_{\text{ord}}}\!\left(u_k - \hat{y}(S, q)\right)^{\!2} \\[3pt] \mathcal{L}_{\text{RPS}} &= \sum_{l=0}^{L-2} \bigl(\text{CDF}_p(l) - \text{CDF}_y(l)\bigr)^2 \end{aligned}$$

Both add a ranked-probability-score term for ordinal questions; Jue-Flash also couples level logits to a regressed expected level.

3. Noul (Boolean) 明辨 · 真伪判词

Calibrated binary truth probability $P(\text{true} \mid S, q) \in [0, 1]$ for rule execution, reachability, prompt-injection detection, and paired counterfactuals.

$$\begin{aligned} P(\text{true} \mid S, q) &= \sigma\!\left(\frac{(z_{\text{true}} - z_{\text{false}}) - b_{\text{noul}}}{T_{\text{noul}}}\right) \\[3pt] \text{conf}(p) &= 1 - H_2(p) \end{aligned}$$

Calibration: Jue uses one global temperature; Jue-Flash uses per-primitive temperatures plus a yes/no bias. Minimal yes/no pairs that need a rule evaluated completely remain the weakest question type for both.

Paired Results · Same Questions

Jue (2.5B) and Jue-Flash (350M) on the Same Benchmark Items

Each row scores both models on exactly the same items with the benchmark's official metric as implemented in Jue's converters. Notes say when a number is not comparable to published leaderboards, when Jue read a sample, and how often Jue-Flash truncated the input. Across the 131 paired tasks, Jue-Flash scores higher on 14.

Engineering & Architecture

Jue-Flash: Mechanisms of a 350M Decision Encoder

FlashDecisionModel combines LiquidAI/LFM2.5-Encoder-350M with the mechanisms below. They are design descriptions; their measured effect is in the paired tables above, where Jue-Flash trails the 2.5B Jue on most external benchmarks.

MECHANISM 01 · HYBRID LFM2.5 + MULTI-DEPTH GATE

LivConv + Query-Conditioned 4-Layer Gate

Combines 10 fast gated short-range convolution (LivConv) blocks and 6 bidirectional group-query attention blocks with a dynamic depth gate over layers (-12, -8, -4, -1), routing syntax to middle layers and policy synthesis to top layers.

MECHANISM 02 · DUAL-PATH READOUT + WATER-FILLING

Global Bilinear + Late-Interaction MaxSim

Fuses Path A (attentive span pooling + bilinear MLP) with Path B (ColBERT-style token MaxSim) and water-filling option packing for wide option lists.

MECHANISM 03 · CHUNKED PACKING & CROSS-ATTN

8-Question Chunking & 2-Hop Cross-Attention

Splits dense multi-predicate requests into 8-question chunks with 2-hop state cross-attention.

MECHANISM 04 · ORDINAL GEOMETRY

Unimodal Ordinal Score Coupler

Couples per-level rubric span logits with a continuous expected-severity regression $\hat{y} \in [0, 1]$ via a learned quadratic distance penalty $-\exp(\gamma_{\text{ord}})(u_k - \hat{y})^2$.

Interactive Wire Protocol Inspector

Recorded POST /v1/systemone Responses

Select a real-world cybersecurity scenario below to inspect how a single packed JSON request containing Choice, Score, and Noul questions is answered simultaneously by Jue and Jue-Flash. These are real responses recorded from both models, not a live service; ✓ / ✗ compare the top answer with the answer the scenario was written for, and both models get several wrong.

POST /v1/systemone — Packed Request Payload JSON Wire Protocol 

          


          

            

              Recorded Output Distributions
              Jue vs. Jue-Flash
            

            

          

        

      

    


    
    

      

        

          
Full-Spectrum Coverage

          
Cybersecurity Decision Domains

          

            Cyber defense questions the models are built to answer, in English and Chinese. Coverage differs between the two models, and not every domain has its own external benchmark; see the paired tables for what was measured.
          

        


        

          

            DOMAIN 01
            
SOC Alert Triage & SIEM

            
SIEM incident grading, alert correlation, false-positive suppression, and escalation routing.

          

          

            DOMAIN 02
            
Edge WAF & API Payloads

            
SQLi, XSS, SSRF, Deserialization, and BOLA/BFLA API anomaly adjudication.

          

          

            DOMAIN 03
            
EDR Process Trees & Sysmon

            
GUID process-lineage graph traversal, LOLBins command-line deobfuscation, and PowerShell/Bash script blocking.

          

          

            DOMAIN 04
            
Network & Firewall (NSM)

            
iptables/nftables/AWS SG/NACL first-match rule evaluation, Zeek/Suricata flow triage, and OT Modbus/DNP3 allowlists.

          

          

            DOMAIN 05
            
Detection Engineering

            
Sigma rule evaluation, YARA rule matching, SQLite threat-hunt query verification, and fail2ban windowing.

          

          

            DOMAIN 06
            
SAST & Code Vulnerability

            
Static-analysis rule matching and security-patch decisions. Repository-scale bug finding was a goal of Jue and was not reached.

          

          

            DOMAIN 07
            
SCA & Supply-Chain Security

            
CycloneDX/SPDX SBOM dependency graphs, OpenVEX/CSAF reachability, semver resolution, and SLSA/Sigstore provenance.

          

          

            DOMAIN 08
            
Cloud IAM & Identity

            
AWS IAM policy evaluation with conditions, Active Directory SDDL access checks, Entra Conditional Access, and GitHub OIDC.

          

          

            DOMAIN 09
            
Kubernetes & Containers

            
Pod Security Admission (PSS), K8s RBAC & NetworkPolicy, OPA/Kyverno admission control, and multi-stage Dockerfiles.

          

          

            DOMAIN 10
            
Vulnerability Management

            
CVSS v3.1 metric prediction, CISA SSVC decision trees, bilingual CVE triage, and KEV prioritization.

          

          

            DOMAIN 11
            
Cyber Threat Intelligence (CTI)

            
MITRE ATT&CK Enterprise/Mobile/ICS technique mapping, D3FEND countermeasures, and CWE root-cause mapping.

          

          

            DOMAIN 12
            
Email & Phishing Defense

            
SPF, DKIM, DMARC alignment evaluation, header spoofing detection, and BEC social-engineering intent classification.

          

          

            DOMAIN 13
            
Host Hardening & CIS

            
CIS-style hardening checks, compliance baselines, sudoers/PAM/cron privilege checks, and pg_hba.

          

          

            DOMAIN 14
            
DLP & Secret Triage

            
High-entropy credential vs. test-fixture disambiguation, PII/PHI exposure scoring, and git commit secret gating.

          

          

            DOMAIN 15
            
Global Cyber Law & Standards

            
Incident-reporting deadlines and obligations across NIS2, DORA, GDPR, HIPAA, SEC 8-K, NIST 800-53, and Chinese cybersecurity law.

          

          

            DOMAIN 16
            
AI Agent Tool Guardrails

            
Real-time pre-execution gating of MCP/function calls, indirect prompt-injection blocking, and least-privilege parameter auditing.

          

        

      

    

    
    

      

        

          
Download · Cite

          
Weights, Code and Report

          

            Each repository holds the weights, a calibration file, a small /v1/systemone server and the
            fifteen example requests from its model card.
          

        

        

          

            JUE · 2.5B · APACHE-2.0
            
huggingface.co/Jue1Dev/jue

            
LoRA adapter for openbmb/MiniCPM5-2B, calibration, server, examples, model card in English and Chinese.

          

          

            JUE-FLASH · 350M · LFM OPEN LICENSE v1.0
            
huggingface.co/Jue1Dev/jue-flash

            
Fine-tuned LiquidAI/LFM2.5-Encoder-350M with decision head, calibration, server, examples. Code is Apache-2.0.

          

          

            TECHNICAL REPORT
            
English · 中文

            
Interface, both models, how they were measured, results on the same items, and what does not work.

          

        

        
@misc{jue2026,
  title  = {Jue: Calibrated Decision Models for Cybersecurity},
  author = {{Jue1Dev}},
  year   = {2026},
  note   = {Technical report},
  url    = {https://huggingface.co/Jue1Dev/jue}
}