What happened
On Oct 6 Mistral launched the ML4 public preview: a 1T-parameter natively multimodal model (49B active) trained on 3,800 Grace Blackwell GPUs in European datacenters, weights dropping end of Oct. It ranks top-5 globally on the Artificial Analysis Cyber Index, scores 82% on reproduce-and-patch a real OSS vulnerability (highest of any model; Claude Opus 5.5 / GPT-6 Astra score near zero due to refusals), and 93% on Cybench. It is being red-teamed with cyber defenders and state authorities with reduced moderation, and is explicitly geared to run on-prem/private cloud for auditable, sovereign security operations.
Why it matters
Refusal-based capability gaps in frontier closed models are a genuine security-operations problem (defenders need to prove flaws are real; losing a capability mid-incident is itself a risk). An open-weight model with top-tier cyber benchmarks that orgs can self-host under their own policies is a meaningful, sovereign-control option for SOCs, malware analysis, detection-rule writing, and vulnerability prioritization.
Applicability
Defensive security teams (malware analysis, vuln research, detection engineering) and sovereign/air-gapped organizations should pilot ML4 via Mistral Studio preview now and re-evaluate when weights drop end of month.