Intern Large Models

Stav AI Act verdict

AI Act 37/100Overall risk · medium

jurisdictionhigh

SHLAB is a Chinese state-funded institution directly established by the Shanghai municipal government. China's National Intelligence Law (Art. 7) can compel any Chinese organisation to cooperate with state intelligence. Unlike CLOUD Act exposure, this has no safe-harbour equivalent and applies structurally to SHLAB regardless of where model weights are deployed. On-premise EU deployment of open weights removes inference-time data exposure but cannot eliminate this organisational-level risk.

governancehigh

SHLAB is directly established and funded by the Shanghai municipal government and classified as a 'New R&D Institute'. There is no independent board, no private shareholding, and governance accountability runs to government authorities. Director Zhou Bowen leads AI safety policy work for China's CnAISDA, deepening the state–lab integration.

compliancehigh

No EU AI Act compliance statement, no GPAI model office registration, no DPA, no GDPR privacy policy, and no EU legal entity identified. GPAI provider obligations under Articles 53–55 have been in force since August 2, 2025, and full enforcement activates August 2, 2026. SHLAB has not signed the GPAI Code of Practice. EU enterprises deploying InternLM models in regulated contexts carry the full provider compliance burden.

privacymedium

No training data opt-out mechanism, consent framework, or dataset copyright clearance documentation published. Training data sourcing described only at a high level (e.g., '4T high-quality tokens'). This raises GDPR Article 10 training-data compliance questions and EU copyright law concerns for EU deployers fine-tuning on InternLM weights.

jurisdictionmedium

SenseTime — the entity credited with 'equal contribution' on the original 2023 InternLM technical report — is on the US Entity List for Xinjiang surveillance abuses. This historical co-development link does not impose a current sanction on SHLAB or current InternLM3 models, but is supply-chain due-diligence relevant for regulated-sector customers, particularly those screening for US-nexus compliance.

securitymedium

No published security certifications (SOC 2, ISO 27001), no formal bug bounty programme, and no responsible disclosure policy identified. Security posture for SHLAB's model release infrastructure and the intern-ai.org.cn chat/API endpoint is opaque to external audit.

compliancemedium

Intern-S1-Pro (released February 2026) has 1T total parameters via MoE architecture. Training compute (FLOPs) has not been disclosed. If training compute exceeds 10^25 FLOPs, Intern-S1-Pro would be classified as a GPAI model with systemic risk under EU AI Act Article 51, triggering enhanced obligations (Article 55) that EU deployers would need to satisfy themselves absent SHLAB engagement.

compliancelow

InternLM2-series model weights require completion of a commercial licence application form. Enterprises must document completion of this form to demonstrate licence compliance; failure to do so creates an IP compliance gap. InternLM3, InternVL3/3.5, and Intern-S1-Pro are fully permissive (Apache 2.0 / MIT) with no application step.

Stav AI Act assessment

Editorial assessment, not legal advice. Stav's risk ratings, scores, and verdicts are our own analysis of publicly available information and may be incomplete or out of date. Verify independently before making compliance or procurement decisions.

Safeguards in place

transparencyUnknown
Published detailed technical reports for InternLM (2023, GitHub), InternLM2 (arXiv 2403.17297, March 2024), and Intern-S1-Pro (2026), with model architecture, training approach, evaluation results across 30+ benchmarks, and acknowledged limitations including potential for harmful outputs.
transparencyUnknown
InternLM3 code and weights are licensed under Apache 2.0 — a fully permissive, well-understood open-source licence with no application step. InternVL3/3.5 weights are under MIT licence. Both make IP compliance clear for EU enterprises.
communityUnknown
Active open-source community with 46+ GitHub repositories, full-chain toolchain (InternEvo, LMDeploy, XTuner, OpenCompass, MinerU) published under Apache 2.0, and hundreds of thousands of developers engaged. Multiple papers accepted at ICLR 2026 (STAR-Bench, CapRL), CVPR, and NeurIPS.
complianceUnknown
SHLAB published AI safety benchmarks evaluating 18 LLMs across frontier AI risk areas at WAIC 2025 (July 2025), and has previously published benchmarks covering value alignment, adversarial robustness, and toxic content risk — indicating genuine engagement with safety research, even within the domestic Chinese framework.
stabilityUnknown
Consistent cadence of major model releases (InternLM → InternLM2 → InternLM2.5 → InternLM3 → InternVL3 → InternVL3.5 → Intern-S1-Pro → Intern-S2-Preview) through 2023–2026, backed by state-guaranteed funding, indicating stable long-term model maintenance capacity.
transparencyUnknown
Full self-hostable open-source model development toolchain (InternEvo for training, LMDeploy for deployment, XTuner for fine-tuning, OpenCompass for evaluation) gives EU deployers full auditability of the end-to-end stack and eliminates call-home dependencies when deployed on-premise.

Published safeguards & certifications

Privacy posture

Flagged issues: 1

No training data opt-out mechanism, consent framework, or dataset copyright clearance documentation published. Training data sourcing described only at a high level (e.g., '4T high-quality tokens'). This raises GDPR Article 10 training-data compliance questions and EU copyright law concerns for EU deployers fine-tuning on InternLM weights.medium

Privacy policy review

Identity & profile

Country: 🇨🇳 China
Open-weights policy: Mixed
Models in catalogue: 0
Website: https://internlm.intern-ai.org.cn
GitHub: https://github.com/InternLM
HuggingFace: internlm · 1,004 followers
HF published: 115 models · 19 datasets · 16 papers

Creator profile

About Intern Large Models

InternLM is the open-weights model series produced by Shanghai AI Laboratory (SHLAB), a research institute established in 2020 by the Shanghai municipal government and operating under state funding — making it structurally accountable to Chinese government authorities rather than to independent shareholders. Models are released under Apache 2.0 for code, with weights freely available for research and commercial use subject to a separate application, and the lab actively publishes technical reports and benchmarks. EU regulated-sector customers should weigh the state affiliation and the absence of GDPR-aligned data processing agreements, EU legal entities, or published security certifications against the lab's genuine open-weights transparency posture and active AI safety research engagement.

Stav editorial summary

Stav’s creator score.