ScamGuard 1.7B 1.7B
flowxai/scam-guard-qwen17b
Scam verdict — macro-F1 · ScamGuardBench v0.2
120-item in-distribution slice. The 1.7B on-device model beats a frontier API on verdict and tactic F1, with zero false positives on legitimate messages.
Research · Open models
Small open models,frontier where it counts.
Apache-2.0 models fine-tuned for the narrow, high-stakes jobs inside regulated workflows — scam detection, PII redaction, risky-clause flagging. Small enough to run on-device, benchmarked head-to-head against frontier APIs.
apache-2.0 · on-device · benchmarked
Industry models
Model families for regulated verticals.
On-device NER and de-identification families, one per industry — small encoder and document models that read your paperwork, extract the entities that matter, and strip regulated data before a single byte leaves your network.
- 9 modelsBanking
OpenLedger
Open-source financial NER
Small encoder models for banking: extract IBANs, PANs, counterparties and 30+ financial entity types, and de-identify PCI & PII — without a single byte leaving your network. IBAN/card checksums validated on-device.
- Extracts
- 30+ financial entities
- Latency
- ~140 ms / doc
- Compliance
- PCI · GLBA · GDPR · AML
github.com/flowx-ai/opennerExplore OpenLedger → - 8 modelsInsurance
OpenCover
Open-source insurance NER
A family of encoder models that read dec pages, FNOL notices, ACORD forms, adjuster notes and medical reports — extracting 28+ entity types and stripping PII + all 18 HIPAA identifiers on-device.
- Extracts
- 28+ insurance entities
- Latency
- ~150 ms / doc
- Compliance
- GLBA + HIPAA
github.com/flowx-ai/opennerExplore OpenCover → - 8 modelsTrade & shipping
OpenFreight
Open-source logistics NER
Encoder and document models that read B/Ls, air waybills, commercial invoices, packing lists and customs declarations — validating ISO 6346, IMO, HS and UN/LOCODE codes and de-identifying pricing & PII on-prem. ONNX/edge exports included.
- Extracts
- 30+ trade entities
- Latency
- ~150 ms / doc
- Compliance
- On-prem · signed audit trail
github.com/flowx-ai/opennerExplore OpenFreight → - 5 modelsPharmacy & healthcare
OpenVita
On-device healthcare NER
On-device NER and SLM models for pharmacy and healthcare ops: route the inbox, extract every expiry, lot and GTIN, and de-identify patient PII and health data — because patient data cannot go to a cloud LLM.
- Extracts
- Expiry · lot · GTIN · PHI
- Latency
- ~126 ms / doc
- Compliance
- HIPAA + GDPR Art. 9
github.com/flowx-ai/opennerExplore OpenVita →
- 1.7BScam & fraud detection
ScamGuard 1.7B
flowxai/scam-guard-qwen17b
On-device scam and fraud detector for SMS, email and chat (English + Romanian). Explains why a message is risky and routes to trusted verification — no network calls.
- Base
- Qwen3-1.7B (LoRA)
- License
- Apache-2.0
- 0.6BScam & fraud detection
ScamGuard 0.6B
flowxai/scam-guard-qwen06b
The phone-sized sibling of ScamGuard: the same on-device scam triage in 0.6B parameters, for the tightest edge budgets.
- Base
- Qwen3-0.6B (LoRA)
- License
- Apache-2.0
- ~300MPII / sensitive-entity detection
CEE-PII
flowxai/cee-pii
Small multilingual span-level PII detector weighted toward Central & Eastern European languages. Redacts before text leaves the perimeter or reaches an LLM — on consumer CPU.
- Base
- GLiNER multi-v2.1 (mDeBERTa-v3)
- License
- Apache-2.0
- 4BRisky-clause detection
Caveat
flowxai/caveat
On-device model that spots and explains risky clauses in consumer contracts across 34 clause types (EN/RO/PL/HU) — privately, with no external APIs. A triage aid, not legal advice.
- Base
- Qwen3-4B
- License
- Apache-2.0
- 4BEscalation gate
Sentinel Gate
flowxai/sentinel-gate
An escalation gate for regulated decisions: decides which cases are safe to automate and which must route to a human, with a structured rationale and audit trail.
- Base
- Qwen3-4B
- License
- Apache-2.0
- 4BOntology extraction
Semantic Mapper
flowxai/semantic-mapper
First-stage extraction for compliance pipelines: turns regulatory clauses into structured ontology JSON (structural, semantic and governance facets) for downstream policy and escalation models.
- Base
- Qwen3-4B (LoRA)
- License
- Apache-2.0
Benchmarks
Measured against the frontier.
Every number below is transcribed from the model card on Hugging Face. Scores are self-reported on held-out sets; frontier comparisons use the slice each card published. The point isn't to win every metric — it's to get frontier-class results on the specific job while running open and on-device.
vs frontier
ScamGuard 0.6B 0.6B
flowxai/scam-guard-qwen06b
Scam verdict — macro-F1 · ScamGuardBench v0.2
Even at 0.6B it edges a frontier API on the in-distribution benchmark with zero false positives; on fresh out-of-distribution messages the frontier model leads.
CEE-PII ~300M
flowxai/cee-pii
PII detection — micro-F1 (exact match)
Held-out multilingual set (EN/PL/RO/HU/UZ); 100-doc slice vs Claude. Fine-tuning lifts exact-F1 4.7× over the zero-shot base, reaching ~89% of a frontier API’s score while running fully offline.
Caveat 4B
flowxai/caveat
Risky-clause — overall F1 · RedFlag-Bench v0.1
100-chunk slice vs Claude Opus. The 4B on-device model trails frontier overall, but leads on the most common clause types (lease 0.80 vs 0.30, English 0.63 vs 0.27) while running fully private.
Sentinel Gate 4B
flowxai/sentinel-gate
Escalation gate — held-out (n=71)
Self-reported on 71 realistic-synthetic regulated cases; no frontier baseline published. Every case needing escalation was flagged — zero missed escalations.
Semantic Mapper 4B
flowxai/semantic-mapper
Ontology extraction — held-out (112 multilingual docs)
Self-reported; no frontier baseline. Perfect structural validity; concept-F1 reflects agreement with FlowX’s annotation convention on unseen documents, not human-legal agreement.
FlowX modelFrontier APIBaseline
Get the weights
Open weights,Apache-2.0.
Pull any model from Hugging Face and run it in your own perimeter — or ask us to fine-tune one for your data.