Maritaca AI AI Vendor Risk Profile
Brazilian foundation-model startup spun out of Unicamp's AI research group. Builds the Sabiá family of Portuguese-language LLMs (Sabiá-2, Sabiá-3) which outperform GPT-4-turbo on Brazilian language benchmarks. Privacy-first API: inputs are discarded immediately after response and not used for training.
Risk overview
Risk score: 49/100
Risk tier: Elevated
Safety rating: 51/100 (higher is better)
Lower risk scores indicate lower assessed risk.
Last verified: May 15, 2026 Fresh
Risk dimensions
| Dimension | Risk score |
|---|---|
| Data Handling | 58/100 |
| IP Exposure | 26/100 |
| Jurisdiction | 29/100 |
| Security | 70/100 |
| Regulatory Compliance | 80/100 |
| Transparency | 45/100 |
| Business Stability | 52/100 |
| Dependency Chain | Not assessed |
| Agent Governance | Not assessed |
Analyst summary
Rating: Acceptable
Maritaca is the only credible Brazilian frontier-builder. Founded in 2022 out of Unicamp's AI research nucleus, the Sabiá model family is purpose-built for Portuguese and beats GPT-4-turbo on 23 of 64 Brazilian-context benchmarks. The privacy story is the genuine differentiator: API inputs are discarded immediately after response generation and explicitly excluded from training, with LGPD-aligned PII handling. Funding is light ($1M Google plus undisclosed follow-on) and the team is small (11-50), so this is a bet on a research-led specialist rather than a scaled enterprise vendor.
Bottom line: The right choice when the answer to "do we need real Portuguese fluency and Brazilian context" is yes. Compliance posture is informal but the privacy default is genuinely stronger than US frontier vendors. Treat as a specialist for Brazilian deployments rather than a general-purpose substitute for OpenAI or Anthropic.
Strengths
- Privacy-default: API inputs discarded after response, never used for training — explicitly documented
- LGPD-compliant Brazilian data residency
- Sabiá-2 and Sabiá-3 outperform GPT-4-turbo on Brazilian-specific benchmarks (legal, education, institutional)
- Academic research lineage (Unicamp NLP group) — Sabiá research is published and reproducible
- Only credible domestic option for Portuguese-language Brazilian deployments
Concerns
- Light funding (~$1M disclosed) and small team — concentration risk
- No SOC 2, ISO 27001, or HIPAA attestation
- GDPR / CCPA compliance not formally claimed
- Public subprocessor list not available
- Bug-bounty and formal red-teaming programs not in place
- Single founder-led startup — succession and continuity risk
Best for
- Brazilian enterprises with Portuguese-language workloads and LGPD residency requirements
- Legal, education, and public-sector buyers in Brazil
- Privacy-sensitive teams who value a hard "no training" guarantee over enterprise-feature breadth
Avoid if
- You need a vendor with formal SOC 2 or ISO certification
- You require HIPAA BAA for protected health information
- Your workload requires a 1000+ FTE vendor with global support coverage
Citations
- Data Handling — Input Logging Policy
- Data Handling — Trains On User Data
- Governance — Research Publications
- Vendors — Founded Year
- Vendors — Description