Loading the catalogue…
Loading the catalogue…
Anthracite is an informal, pseudonymous online collective with no disclosed headquarters, legal entity, or named leadership — it operates via HuggingFace and Open Collective crowdfunding, so its jurisdiction and CLOUD Act exposure cannot be positively verified. Its Magnum model series are openly released fine-tunes on top of third-party base models (Qwen, Mistral, Gemma, Yi) using training data that includes Claude-output-derived datasets, which raises a chain-of-title/IP concern given Anthropic's terms prohibit training competing models on Claude outputs. There is no privacy policy, DPA counterparty, EU AI Act compliance documentation, or security certification of any kind, making the creator unsuitable as a default choice for regulated-sector EU deployments outside experimental or fully air-gapped, self-hosted use.
Anthracite is an anonymous, unincorporated volunteer collective with no legal entity, named leadership, registered address, or determinable jurisdiction — there is no legal counterparty for a DPA and no liability chain for enterprise deployment.
No privacy policy, GDPR documentation, DPA template, DPO contact, or any data protection infrastructure exists; GDPR compliance is entirely unverifiable.
Magnum training datasets include Claude-output-derived data, and the series' explicit design goal is to replicate Claude 3 Sonnet/Opus prose — creating potential IP/ToS chain-of-title risk given Anthropic's usage policies prohibit training competing models on Claude outputs.
No EU AI Act compliance statement, no Article 53 GPAI transparency documentation, no formal training data summary meeting regulatory standards, and no Code of Practice engagement.
Jurisdiction is entirely unverifiable due to pseudonymous contributors; the current CLOUD Act exposure value of 'false' reflects absence of evidence rather than confirmed non-US status.
No security certifications (SOC 2, ISO 27001), bug bounty programme, or responsible disclosure policy exists; security posture is entirely opaque.
Stav’s 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risk ratings, scores, and verdicts are our own analysis of publicly available information and may be incomplete or out of date.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.
Training datasets and full training configuration details are publicly disclosed in model card READMEs, giving more training-side visibility than many proprietary labs, even though this also surfaces the IP concern.
Financial transparency via Open Collective: donations and expenditures are publicly visible, with no hidden corporate funding or undisclosed investors.
Models are released with full downloadable weights across multiple versions and formats (including quantised GGUF/exl2), enabling independent evaluation and fully local, air-gapped deployment without contacting Anthracite infrastructure.
Consistent iterative releases (v1 through v4 series across multiple base models) show sustained development activity from at least August 2024 through October 2024, with continued third-party tracking into 2026.
Published safeguards & certifications