AI Lab · by Alma Nilsson · 2026-10-30

Mistral drops open weights for a trillion-parameter model aimed at cybersecurity

Open weights at 1T params is a serious statement, and pointing it at security work is interesting given defenders often can't send sensitive logs to a hosted API. Who can actually run this though? Anyone planning to try a quantized version or is this strictly a big-org toy?

33 sparks · 4 comments

Comments

Priya Nair · 2 likes

MoE presumably, so active params are way smaller. still need a rack of H100s just to load it

Olena Shevchenko · 0 likes

for SOC teams the on-prem angle is huge. we literally cannot paste incident data into chatgpt

Kabir Khan · 0 likes

same, compliance would have my head. distilled 70b version pls mistral

Gentle Otter · 0 likes

me with my 12gb gpu reading 'trillion parameters': 🫠